Add native Perfetto (.pftrace) export to torch.profiler

[ajassani]

Summary

Adds a native Perfetto (.pftrace) export path to torch.profiler, so users can write Perfetto protobuf traces directly from a profiling session instead of exporting Chrome JSON and converting afterwards.

• New export_pftrace(path) on torch.profiler.profile (_KinetoProfile) and on the legacy torch.autograd.profiler.profile, mirroring export_chrome_trace (including .gz handling).
• C++ bindings: ProfilerResult::savePftrace / ActivityTraceWrapper::savePftrace, routed through Kineto's pftrace:// save protocol, exposed to Python as save_pftrace.
• Build gated behind a new USE_KINETO_PERFETTO CMake option (default OFF), plumbed to Kineto's KINETO_ENABLE_PERFETTO.

When Kineto is built without Perfetto support, behavior is unchanged and the existing JSON path is untouched.

This is the PyTorch-side companion to the Kineto change that implements the actual Perfetto trace logger and the pftrace:// protocol registration.

Usage

with torch.profiler.profile(activities=[...]) as prof:
    run_workload()
prof.export_pftrace("trace.pftrace")

Test plan

• Builds with USE_KINETO_PERFETTO=ON against a Kineto built with Perfetto support (MI300-class GPU, ROCm).
• CPU-only smoke workload produces a non-empty .pftrace that loads in the Perfetto UI.
• GPU smoke workload produces a non-empty .pftrace with device track events.
• CI build with default USE_KINETO_PERFETTO=OFF (JSON path unchanged).