feat: add CUDA graph debug dumps

[popsiclexu]

Summary

Adds an optional CUDA graph debug dump path for MUSA graph capture:

• Read TORCHADA_CUDA_GRAPH_DEBUG_DUMP_PATH once during CUDA graph patch setup and cache the dump directory.
• Call graph.enable_debug_mode() before graph capture when the dump directory is configured.
• Call graph.debug_dump(...) after a successful graph capture and write timestamped files named graph_<time_ns>.dot.
• Document the environment variable in English and Chinese READMEs.
• Add a real MUSA GEMM graph capture test that verifies a dot file is emitted and removes the dump file afterward.

Validation

TORCH_DEVICE_BACKEND_AUTOLOAD=0 PYTHONPATH=src python -m pytest tests/test_cuda_patching.py::TestCUDAGraph -q

Result:

12 passed, 2 warnings in 9.93s

The warnings are from torch_musa debug dumping and confirm the generated dot path:

/tmp/pytest-of-root/pytest-1/test_graph_context_manager_deb0/graph_dumps/graph_1783672385347873460.dot

[augmentcode]

🤖 Augment PR Summary

Summary: Adds optional MUSA/CUDA-graph debug “dot” dumping support to help diagnose captured graph structure.

Changes:

• Introduces TORCHADA_CUDA_GRAPH_DEBUG_DUMP_PATH to enable graph debug dumps and caches the configured dump directory during patch setup.
• Extends the torch.cuda.graph context-manager wrapper to call enable_debug_mode() before capture and debug_dump(...) after a successful capture.
• Generates timestamped graph_<time_ns>.dot filenames to avoid overwriting across repeated captures.
• Documents the new environment variable in both English and Chinese READMEs.
• Adds a MUSA-only GEMM capture test that asserts a dot file is emitted and cleans it up afterward.
• Includes minor test formatting cleanups (path join / imports) with no functional change.

Technical Notes: Debug dumping is opt-in via env var, errors are surfaced as warnings, and the dump path is prepared (created) up-front during patch initialization.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 2 suggestions posted.

Comment augment review to trigger a new review at any time.

[augmentcode]

src/torchada/_patch.py:404: _cuda_graph_debug_dump_dir caches dump_dir exactly as returned by expanduser(), which can still be a relative path (e.g. ./graph_dumps); if the process later changes CWD, dumps may be written to an unexpected location.

Severity: medium

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}

[augmentcode]

src/torchada/_patch.py:506: GraphWrapper doesn’t retain the graph_obj passed into __init__ and instead re-discovers it via musa_graph/cuda_graph attributes on the wrapped context manager; if those internals differ across torch/torch_musa versions, debug enable/dump could silently no-op.

Severity: medium

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}