Rewrite parcagpu on Proton's CUPTI infrastructure with PC sampling

[gnurizen]

Replace the standalone C CUPTI interposer with a C++ implementation
built on Proton's callback and activity APIs. This integrates CUPTI
activity tracing, kernel correlation filtering, and PC sampling with
stall reason analysis into a single shared library.

Key additions:

• src/cupti.cpp: CUPTI activity subscriber using Proton callbacks,
  with USDT probes for activity_batch, cubin lifecycle, stall reasons,
  pc_sample_batch, and error reporting
• src/pc_sampling.cpp: PC sampling with probabilistic windowed
  start/stop, KERNEL_SERIALIZED mode, and token-bucket rate limiting
• src/correlation_filter.cpp: lightweight correlation ID filter for
  matching kernel activities to profiled ranges
• ebpf/cupti_bpf.h: shared BPF-side layouts for CUPTI activity and
  PC sampling records
• test/bpf/: eBPF activity parser test using parca-dev/usdt for USDT
  argument extraction, with cubin loading, SASS lookup, and stall
  reason display
• test/mock_cupti.c, test/test-pc-mock.sh, test/test-pc-real.sh:
  mock and real PC sampling test harnesses
• microbenchmarks/: CUDA micro-benchmarks for PC sampling validation

Build system switched to CMake with Proton as a git submodule.

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com