build: add CMake package config

[voltjia]

Summary

• Install InfiniRTConfig.cmake, InfiniRTConfigVersion.cmake, and the namespaced InfiniRT::infinirt target so downstream CMake projects can use find_package(InfiniRT CONFIG REQUIRED).
• Export relocatable backend dependency metadata and replace the installed-consumer check with a separate CMake project that finds, validates, and links the installed package.
• Update the consumer example and installation documentation to use CMAKE_PREFIX_PATH and the imported target.

Motivation

Installed InfiniRT prefixes currently expose headers and libinfinirt without a CMake package, so each downstream project must rediscover the include directory, library, language requirement, and backend dependencies. A versioned package config provides a standard downstream interface and avoids exporting backend SDK paths from the build machine.

N/A - no linked issue.

Type of Change

• feat - new feature / new backend capability / new public API
• fix - bug fix
• perf - performance improvement without behavior change
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• Hygon (WITH_HYGON)
• MetaX (WITH_METAX)
• Moore (WITH_MOORE)
• Cambricon (WITH_CAMBRICON)
• Ascend (WITH_ASCEND)
• Build system / CMake / generated headers
• Public headers / installed consumer API
• Documentation only

Smoke Build and Test Result

All platform checkouts were pinned to commit 9ae8f8fd0b0ac4807a0d8b325fb0d3b2809fb161. NVIDIA, Iluvatar, Moore, Cambricon, and Ascend builds ran in the accelerator-dev/<platform>:latest images with --privileged. MetaX device smoke used its documented image, while the build ran directly on the same host because that image does not contain CMake; the host provides CMake 3.28.3 and the same /opt/maca SDK.

# CPU-only package and installed-consumer validation
cmake -S . -B build-cpu \
  -DCMAKE_BUILD_TYPE=Release \
  -DWITH_CPU=ON \
  -DINFINI_RT_BUILD_TESTING=ON
cmake --build build-cpu -j 8
ctest --test-dir build-cpu --output-on-failure

100% tests passed, 0 tests failed out of 6

# Common accelerator configure/build/test shape
cmake -S . -B build-pr28 \
  -DCMAKE_BUILD_TYPE=Release \
  -DWITH_CPU=ON \
  -DWITH_<PLATFORM>=ON \
  -DINFINI_RT_BUILD_TESTING=ON
cmake --build build-pr28 -j 8
ctest --test-dir build-pr28 --output-on-failure

# Platform options used
-DWITH_NVIDIA=ON
-DWITH_ILUVATAR=ON
-DWITH_METAX=ON
-DWITH_MOORE=ON
-DWITH_CAMBRICON=ON
-DWITH_ASCEND=ON

Each completed platform test set includes test_install and test_install_consumer. The latter configures an external project with an exact version requirement, resolves the installed package through CMAKE_PREFIX_PATH, validates InfiniRT_ENABLED_BACKENDS and every InfiniRT_WITH_<BACKEND> value, links InfiniRT::infinirt, and runs the resulting executable. A negative configure check also confirmed that mismatched expected metadata is rejected.

Device tensor smoke results and image IDs:

NVIDIA:    cuda:0 [1.0]  sha256:dd94fce2f83a180e2271f02f69713de5360aaf916b01e1f7f54173519fd54efd
Iluvatar:  cuda:0 [1.0]  sha256:5dc3abe5c210b9160cc936f42929f9c5b100fefff175af15e67ce0d359b320ce
MetaX:     cuda:0 [1.0]  sha256:f60d250f4409e4c1a7817dce498905382bc58f1ccc291834e1dc1ecfa3bbf7f5
Moore:     musa:0 [1.0]  sha256:70f9b7c621bc01b69a422501f139699d23086dfe0b61df99cc92a1de07025ce6
Cambricon: mlu:0 [1.0]   sha256:d3544201bc7039f328abf137d154b24a89a17fdd48a487dc0ae0bdd6d14a5cf6
Ascend:    npu:0 [1.0]   sha256:7da2cb1fa09f19d55fa8b41edae49d9a1eda7fd87bdbddf2d6a943f68e6a2f2b

InfiniOps downstream validation used current InfiniTensor/InfiniOps master at 296271487beb594a248fd463e5fff14f7ab74293. Its temporary checkout was changed only to consume InfiniRT::infinirt; configuration received CMAKE_PREFIX_PATH and no legacy InfiniRT root, include, or library variables.

-- Using InfiniRT 0.1.0 CMake package: /workspace/infinirt/build/tests/install_consumer_prefix/lib/cmake/InfiniRT
Successfully built infiniops-0.1.0-cp312-cp312-linux_x86_64.whl
Running 70 items in this shard
56 passed, 14 skipped, 8975 deselected in 7.76s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Yes	full passed	6/6 CPU-only tests passed; also covered in the combined CPU + NVIDIA build.
NVIDIA	Yes	full passed	8/8 InfiniRT tests passed; InfiniOps downstream smoke: 56 passed, 14 skipped.
Iluvatar	Yes	full passed	7/7 tests passed in the documented Iluvatar image.
Hygon	Yes	Blocked	ssh hygon (10.211.3.27:22) timed out from the local host and five accelerator hosts; no container or source command ran.
MetaX	Yes	full passed	7/7 tests passed on the MetaX host with /opt/maca; the documented image supplied the successful device smoke but lacks CMake.
Moore	Yes	full passed	7/7 tests passed in the documented Moore image.
Cambricon	Yes	full passed	7/7 tests passed in the documented Cambricon image.
Ascend	Yes	full passed	7/7 tests passed in the documented Ascend image; the image exhibited the documented post-output exit behavior.

Full `ctest` summary

CPU-only:
100% tests passed, 0 tests failed out of 6
Total Test time (real) = 0.88 sec

NVIDIA:
100% tests passed, 0 tests failed out of 8
Total Test time (real) = 5.73 sec

Iluvatar:
100% tests passed, 0 tests failed out of 7
Total Test time (real) = 0.67 sec

MetaX:
100% tests passed, 0 tests failed out of 7
Total Test time (real) = 1.01 sec

Moore:
100% tests passed, 0 tests failed out of 7
Total Test time (real) = 2.71 sec

Cambricon:
100% tests passed, 0 tests failed out of 7
Total Test time (real) = 2.48 sec

Ascend:
100% tests passed, 0 tests failed out of 7
Total Test time (real) = 2.62 sec

Benchmark / Performance Impact

N/A - this changes CMake package metadata, installation checks, examples, and documentation; runtime implementation and dispatch behavior are unchanged.

Notes for Reviewers

• Package version 0.1.0 matches the project version used by downstream packaging.
• CUDA-like packages rediscover CUDAToolkit; MetaX, Moore, Cambricon, and Ascend packages rediscover their SDK runtime targets from consumer-side hints and environment variables. This keeps build-machine SDK paths out of the installed target export.
• The external consumer compares the complete enabled-backend list and all eight backend flags against the source build, so stale or incomplete package metadata fails during CMake configuration.
• Please focus on the SDK rediscovery rules for non-NVIDIA backends. The PR remains draft because Hygon could not be reached at the SSH endpoint in the local configuration.
• The InfiniOps validation demonstrates the intended migration from manual find_path / find_library logic to find_package(InfiniRT CONFIG REQUIRED) and InfiniRT::infinirt.