Add CLI option to set target-time parameter for batch measurement loop

[oleksandr-pavlyk]

new --batch-target-time CLI option

This PR adds support for --batch-target-time CLI option:

* `--batch-target-time <seconds>`
  * Target accumulated GPU time for batched measurements.
  * Default is 0.5 seconds.
  * `<seconds>` must be finite and positive.
  * Batched measurements continue until both `--min-samples` and the
    accumulated GPU-time target are satisfied, unless `--timeout` is reached
    first.
  * Applies to the most recent `--benchmark`, or all benchmarks if specified
    before any `--benchmark` arguments.

This option controls batch size in batch measurement loop by specifying its target duration.

Batch measurement loop termination condition did not change: both the number of launches executed and the time executing them should exceed thresholds: num_samples >= m_min_samples && total_cuda_time >= m_batch_target_time.

Improved error on unrecognized CLI option

An error handling of unrecognized CLI options is improved:

$ ./build/bin/nvbench.example.cpp17.throughput --stopping-criterion entropy --min-time 1

NVBench encountered an error:

nvbench/option_parser.cu:1180: Error handling option `--min-time 1`:
nvbench/option_parser.cu:141: --min-time is not valid for the active stopping criterion 'entropy'.
  Current criterion 'entropy' accepts: --max-angle, --min-r2.

Change in behavior of batch measurements when no prior cold measurements exist

We used to overwrite the value like so: m_min_time = std::max( std::midpoint(m_min_time, m_timeout), m_min_time * 5);.

In default settings (min-time of 0.5 seconds and timeout of 15 seconds) this would change value to 2.5 seconds.
Given then user may hit this branch by using state::exec(exec_tag::hot, launchable); in the benchmark generator, this behavior is very surprising when --batch-target-time is explicitly specified.

For this reason, and because the option is now explicit, this overwrite has been removed.

Relationship to other work

This PR is an alternative to #424. The advantages of this solutions are:

• --batch-target-time option is global, i.e., it is valid irrespective of which stopping criterion is considered active by option parser.
• it allows controlling batched measurements loop independently from specifying stopping criterion which is only applied to cold/cpu_only measurements loops
• --min-time option continues on as stdrel-criterion parameter.

Closes #408

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• New Features
  • Added configurable batched-measurement targeting via the --batch-target-time CLI option (default 0.5s).
  • Exposed batch target time in C++/Python benchmark/state APIs and option decorators.
  • Included batch_target_time in emitted JSON results.
• Improvements
  • Updated hot-trial batching to stop based on accumulated GPU-time vs batch target time, with more robust batch-size prediction.
  • Refreshed timeout/warning messaging and related “skip time” help to reference batch target time.
• Bug Fixes
  • Added validation to reject non-finite or non-positive values, with expanded parsing/decorator tests.

Walkthrough

Adds configurable batch target time for GPU measurements through C++ and Python APIs, CLI parsing, JSON serialization, hot-measurement scheduling, documentation, and tests.

Changes

Batched measurements

Layer / File(s)	Summary
Batch target configuration contracts nvbench/detail/validate_batch_target_time.cuh, nvbench/benchmark_base.*, nvbench/state.*, nvbench/json_printer.cu	Benchmark and state objects validate, store, expose, initialize, clone, and serialize batch target time.
Hot measurement batch scheduling nvbench/detail/measure_hot.*	Hot measurements predict bounded batch sizes and stop after target CUDA time and minimum sample count are reached.
CLI parsing and documentation nvbench/option_parser.cu, docs/cli_help.md	The CLI parses and validates --batch-target-time, reports active criterion parameters consistently, and documents the option and timeout criteria.
Python API and validation coverage python/src/py_nvbench.cpp, python/cuda/bench/_decorators.py, python/test/test_cuda_bench.py, testing/option_parser.cu	Python bindings and decorators expose batch target time, with tests for APIs, decorators, documentation, valid values, invalid inputs, and criterion-parameter errors.

Assessment against linked issues

Objective	Addressed	Explanation
Provide an explicit batch measurement option for batched minimum-time configuration [#408]	✅
Reject invalid criterion parameters with an error identifying the active criterion and accepted parameters [#408]	✅
Move the batched minimum-time setting out of the stdrel criterion model [#408]	❓	The new --batch-target-time option is modeled independently, but the existing --min-time criterion parameter remains in the CLI model.

---

_{Comment @coderabbitai help to get the list of available commands.}