Rebase onto proton and add PC sampling

libparcagpucupti is now based loosely on the proton library so it delegates to the cupti library and supports 12/13 from one binary.

PC sampling is in alpha state and disabled by default.

Add support for batch kernel updates and add rate limiter

What's Changed

• fix leak by @gnurizen in #4
• Go back to small buffer but let allocator worry about caching by @gnurizen in #5
• Capture runtime and driver API callbacks by @gnurizen in #6
• rate limit by @gnurizen in #7
• Replace interval rate limiter with token bucket algorithm by @gnurizen in #8
• Add correlation ID filter to track traced kernel launches by @gnurizen in #9
• Test improvements with realistic CUDA API stack traces by @gnurizen in #11
• Fix test running/building by @gnurizen in #13
• Batch kernel timing activities over new 'activity_batch' probe by @gnurizen in #14

Full Changelog: v0.1.1...v0.1.2

v0.1.1: Add CUDA 13 Support

What's Changed

• Build fixes for new version scheme by @gnurizen in #3

New Contributors

• @gnurizen made their first contribution in #3

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Initial release.

Provide NVIDIA libcupti injections via USDTs for profilers to intercept and use for profiling. Load via the CUDA_INJECTION64_PATH environment variable.

Works together with parca-agent v0.43.0+.