The user-streams support PR #182971 added _ensure_aoti_stream_helpers_emitted()
which generates CUDA-specific AOTIPerThreadEventCache and AOTIPerThreadStreamCache
helper classes, along with cudaEventRecord/cudaEventSynchronize calls.
These helpers are emitted from multiple code paths in CppWrapperGpu:
1. codegen_stream_info_prologue() — when multiple streams detected
2. _emit_stream_op_inline() — when stream ops like record_event are used
3. codegen_enter_cuda_stream_context() — when switching to non-default stream
4. codegen_exit_cuda_stream_context() — when leaving a non-default stream
On XPU, SYCL in-order queues handle events implicitly. Skip all four paths.
Verified on PVC (driver 1.6.33578) and BMG (Intel Arc Pro B60): D2H copy PASS.
Fixes #189327
Fixes #189326