[WIP] PyTorch ROCTx Observer Implementation

[EvanPhoukong]

Motivation

This branch adds a custom ROCTX observer and wires it to a public API so users can annotate entire training regions with emit_roctx() without relying on Kineto for that behavior.

Technical Details

• Standalone observer (C++): torch/csrc/profiler/standalone/roctx_observer.cpp implements ROCTXThreadLocalState and pushROCTXCallbacks(), which register enter/exit callbacks. On activation, the observer emits roctxMarkA("PyTorch_ROCTX_observer_v2_active") so traces can confirm the v2 path; each op pushes/pops a named ROCTX range (with optional input-shape reporting).
• Profiler integration: torch/csrc/autograd/profiler_kineto.cpp routes ProfilerState::ROCTX to pushROCTXCallbacks() (same class of “non-Kineto” profilers as NVTX). ProfilerState / ActiveProfilerType expose ROCTX in the C++ profiler API and Python bindings (torch/csrc/profiler/python/init.cpp, torch/csrc/profiler/orchestration/observer.h).
• Python API: torch.autograd.profiler.emit_roctx (torch/autograd/profiler.py) is the context manager that enables ProfilerState.ROCTX for the wrapped block.
• Explicit ROCTX from Python: Low-level bindings live under torch._C._roctx (torch/csrc/cuda/shared/roctx.cpp) with a torch.cuda.roctx facade (torch/cuda/roctx.py) for manual range / mark style annotations (orthogonal to enabling the autograd observer).

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[rocm-repo-management-api]

Jenkins build for 5b031ede28630bc88ee00467913316c1782cb80a commit finished as FAILURE
Links: Pipeline Overview / Build artifacts / Test Results