Run the ExecuTorch TensorRT delegate on a caller-selected CUDA stream (green-context support)

[shoumikhin]

Summary

The ExecuTorch TensorRT delegate created and owned a private CUDA stream and ran every enqueueV3() on it, so an application could not place inference on a specific CUDA stream or context — in particular a CUDA green context for SM partitioning.

This lets the caller select the stream, giving the libtorch-free ExecuTorch runtime the same caller-stream capability the libtorch TensorRT runtime gained in #4232.

Changes

• Add a scoped CudaStreamGuard (mirroring c10::cuda::CUDAStreamGuard) to select, per calling thread, the CUDA stream the delegate runs TensorRT on. With no guard active the delegate runs on cudaStreamPerThread.
• execute() runs enqueueV3() and the staging copies on the selected stream; init() no longer creates a stream (the delegate owns none).
• Green context: scope a guard with a stream created on the green context via cuGreenCtxStreamCreate; the partition confinement travels with the stream, so the green context need not be made current. cudaStreamPerThread is invalid while a green context is current (cudaErrorInvalidResourceHandle), so a green-context caller must scope a guard.
• cudaSetDevice() is applied only when the engine's device differs from the current device and is restored on exit, so it no longer clobbers a context the caller established.
• Backward compatible: device-resident outputs are left enqueued (no end sync) only while a guard is active; the default path and host-staged outputs still synchronize before returning, preserving the prior "results ready on return" behavior.

Validation

Verified on an H100 (CUDA 12.8) with an %smid probe: a cuGreenCtxStreamCreate stream confines kernels to the green context's SM partition even when the primary context is current; cudaStreamPerThread errors with cudaErrorInvalidResourceHandle while a green context is current; the non-green default path uses the full device.

No dependency on the libtorch Torch-TensorRT runtime or libtorch is added.

Follow-up: a unit test for the stream selection (guarded vs. default) can be added.