fix(torch): honor handle streams in generated operators

[voltjia]

Summary

• Installs the native stream stored in an InfiniOps Handle as PyTorch's current stream before generated ATen and linked Torch calls, then restores the previous device and stream.
• Declares the shared C10<kDev> template in src/torch/c10.h and keeps each backend specialization in src/torch/<platform>/c10.h, following the existing platform-oriented C10 design without a separate stream bridge.
• Makes generated wrappers, linked adapters, and the hand-written FlashAttention wrapper use the same platform C10 interface, while preserving PyTorch's current stream when the handle stream is null.
• Adds the required vendor Torch headers, libraries, runtime paths, generator assertions, and current-stream versus explicit-handle regression coverage.

Motivation

Generated ATen backends received the InfiniOps handle stream through stream_, but they invoked PyTorch without installing that stream as the framework's current stream. PyTorch work could therefore run on a different stream from the caller-selected InfiniRT stream and violate the caller's ordering assumptions. Linked Torch adapters already modeled device-specific C10 behavior, so the shared implementation should retain that template structure and live under the common Torch backend rather than under linked.

No issue is linked.

Type of Change

• feat - New feature, operator, or platform.
• fix - Bug fix.
• perf - Performance improvement without a behavioral change.
• refactor - Code restructuring without a behavioral 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.

Platforms Affected

• CPU (WITH_CPU).
• NVIDIA (WITH_NVIDIA).
• Iluvatar (WITH_ILUVATAR).
• MetaX (WITH_METAX).
• Cambricon (WITH_CAMBRICON).
• Moore (WITH_MOORE).
• Ascend (WITH_ASCEND).
• PyTorch C++ bindings (WITH_TORCH).
• Build system, CMake, or CI.
• Python bindings or user-facing API.

Smoke Test Result

Final validation SHA: f6e6f270df79fce687570ea1b42cbb58fe0ebc9b.

Local NVIDIA generated-plus-linked smoke build:
85 passed, 22 skipped in 9.77s.

Generator and architecture tests:
45 passed in 1.65s.

Generated and linked stream regressions:
4 passed in 2.47s.

GitHub Actions affected-platform matrices:
The main and shadow matrices passed for Iluvatar, MetaX, Cambricon, Moore, and Ascend. The final-SHA NVIDIA main and shadow jobs failed after their assigned runners went offline during execution. Neither job produced a failure log; the main job remained in the host-device unit-test step and the shadow job remained in the remote-task wait step.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Smoke passed locally.	The generated-plus-linked build passed. Generated abs current-stream and handle-stream cases passed, and generated plus linked FlashAttention handle-stream cases passed for implementation slots 8 and 16. The preceding SHA with identical executable code passed the NVIDIA main matrix. The final-SHA main and shadow jobs failed after their assigned runners went offline during execution, without producing failure logs.
Iluvatar	Yes	Main and shadow CI passed.	The CUDA-compatible C10 specialization uses InfiniRT's native stream type and c10::cuda::getStreamFromExternal.
MetaX	Yes	Main and shadow CI passed.	The CUDA-compatible C10 specialization uses InfiniRT's native stream type and c10::cuda::getStreamFromExternal.
Cambricon	Yes	Main and shadow CI passed.	The C10 specialization uses torch_mlu::getStreamFromExternal.
Moore	Yes	Main and shadow CI passed.	The C10 specialization uses c10::musa::getStreamFromExternal from the explicitly linked libmusa_python.
Ascend	Yes	Main and shadow CI passed.	The C10 specialization uses c10_npu::getStreamFromExternal when the installed headers provide NPUStreamUtils.h, and otherwise retains the current stream.

Local validation output.

Ruff 0.15.22 check: passed.
Ruff 0.15.22 format --check: passed.
Clang Format 21: passed.
Generator and architecture tests: 45 passed.
NVIDIA generated stream regression: 2 passed.
NVIDIA generated and linked FlashAttention stream regression: 2 passed.
NVIDIA smoke suite: 85 passed, 22 skipped.

Benchmark / Performance Impact

N/A. This is a stream-ordering correctness fix and an ownership refactor, and no performance benchmark was run.

Notes for Reviewers

InfiniRT remains the source of truth for each backend's native stream type through Runtime<kDev>::Stream. Vendor PyTorch APIs are used only by the corresponding C10<kDev> specialization to install that native stream into PyTorch's thread-local current-stream state.

src/torch/c10.h contains only the template declaration. Each platform owns its specialization under src/torch/<platform>/c10.h, matching the original C10 organization. The old src/linked/torch/c10.h, backend-local linked C10 headers, and the centralized src/torch/stream_.h bridge have been removed.

Generated and linked call sites hold an optional platform StreamGuard. They call C10<kDev>::GetStreamFromExternal only when InfiniRT supplies a non-null stream, so a null handle leaves PyTorch's selected current stream unchanged. The only platform-selection preprocessor conditions are the existing build guards needed to avoid including unavailable vendor headers; stream conversion itself is implemented by template specializations rather than macros.

NVIDIA, Iluvatar, and MetaX use c10::cuda::getStreamFromExternal. Moore uses c10::musa::getStreamFromExternal, Cambricon uses torch_mlu::getStreamFromExternal, and Ascend uses c10_npu::getStreamFromExternal when available.

Hygon is not claimed by this change because a validated PyTorch external-stream bridge for that backend is not available in the current test matrix. Its C10 specialization explicitly retains the current PyTorch stream.