fix(executorch): initialize a CUDA context for default-stream execution

[shoumikhin]

Running a TensorRT program through ExecuTorch can fail on a worker thread that has not used CUDA yet, even when it works on the loading thread. TensorRT 11.3 looks up default-stream contexts through the CUDA driver, and that lookup needs a current context.

This affects engines loaded with shared activation scratch disabled and engines loaded with it enabled that need no scratch. Both skip the pool's capture check, which otherwise makes a context current. With no caller stream, cudaStreamLegacy, or cudaStreamPerThread, their first enqueue can fail.

For a default stream, execute() calls cudaFree(nullptr) after capture refusal and input/output validation, before claiming scratch or enqueueing. This makes the primary context current only when no context is current, preserving a context the caller set. Explicit nullptr gets the call too; caller-created streams are unchanged. A failed call logs CUDA's error, clears it, and returns Error::Internal before enqueueing. Sticky device faults can survive the clear.

Capture remains unsupported. Calls on any stream must not overlap a Global capture on any thread or a ThreadLocal capture on the calling thread. This includes option-on engines that need no scratch. A ThreadLocal capture on another thread is unaffected. The header also clarifies that NotSupported can mean a refused output resize, not just a pool capture refusal.

One existing limit remains: a fresh thread can reject an input or output mapped from another GPU because pointer validation runs before context setup. Such buffers still need the engine's context made current on that thread before execution.

Tests cover each default-stream selection from fresh threads, including option-on scratch-free engines and pooled engines, caller-context preservation, capture refusal, and rejection of bad input. On Linux x86_64 with TensorRT 11.3, the shared scratch backend test file passes 36 of 36, and the new scratch-free test fails three times out of three when only that path is skipped. The explicit nullptr row also passes without this fix. With TensorRT 11.1, all 36 pass, and the fresh-thread regression does not reproduce.