Preserve the CUDA default stream in DSA wrappers

[kunlunl]

What

Use torch.cuda.get_stream_from_external() when entering a caller-supplied CUDA stream.

Why

torch.cuda.ExternalStream(0) does not preserve the legacy default-stream handle. It produces a different pooled stream, so PyTorch setup or copy operations performed inside the wrapper can race a CuTe kernel launched on stream 0.

get_stream_from_external() preserves both the zero default-stream handle and nonzero external stream handles.

Validation

Run on an H100 (SM90) with PyTorch 2.12, CUDA 13.2, and cuDNN Frontend built from current develop:

• New stream-identity regression on unpatched develop: failed as expected because the active handle was nonzero.
• Same regression with this change: 1 passed.
• Complete test/python/fe_api/dsa suite: 22 passed.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 42e7fb12-345d-4ef0-a97a-a957983ac8fb

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[kunlunl]

Superseded by #354, which now contains this fix as a separate signed-off commit and passed the combined H100 DSA validation (3 targeted regressions and 24/24 DSA tests).