raw_launch.rs: RawArgs fixed-array u64 storage + raw_launch() calling cuLaunchKernel directly. Bypasses PushKernelArg trait dispatch, Arc bookkeeping, event guards. Foundation for 500+ sps hot path. vendor/cudarc: added CudaFunction::cu_function() accessor mirroring CudaStream::cu_stream() pattern. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>