[Profiler] Add opt-in ROCTx ranges to record_function

[EvanPhoukong]

Rocprofiler-sdk PR: ROCm/rocm-systems#9109

Motivation

Allow ROCm profiling tools to capture Python record_function annotations as ROCTx ranges.

Technical Details

• Add opt-in ROCTx emission through TORCH_PROFILER_EMIT_ROCTX.
• Lazily import roctx and report a clear error when unavailable.
• Keep ranges balanced across nesting, exceptions, and deferred callbacks.
• Avoid duplicate markers while the NVTX profiler is active.
• Add tests covering lifecycle, configuration, imports, and failures.

Test Plan

1. Run the targeted rocprofv3 option and environment-propagation tests.

2. Build rocprof-sys and run its CLI help and environment-propagation tests.

3. Profile a real ROCm PyTorch workload containing:

   with record_function("optimizer_step"):
       optimizer.step()

4. Verify that both profiler outputs contain an optimizer_step marker region with GPU work inside it.

Test Result

• All targeted rocprofv3 integration tests passed.
• rocprof-sys built successfully and its targeted tests passed.
• Validated on an AMD Instinct MI300X with PyTorch 2.10.0 and ROCm 7.1.
• rocprofv3 and rocprof-sys both captured the optimizer_step range.
• Each captured range enclosed the expected optimizer GPU kernel.