GPU kernel execution profiling vs current gpu.cuda.* metrics

[123liuziming]

OBI already emits CUDA metrics from libcudart uprobes (gpu.cuda.kernel.launch.calls, graph launches, grid/block, malloc, memcpy). For GPU incident response we need Nsight Systems-like kernel execution data (GPU-side duration, stream, kernel name). nsys profile generally has to start the process, so it cannot be turned on for a live job.

That does not look like a fit for more gpu.cuda.* metric attributes: CUDA launches are asynchronous, so uprobe timestamps are enqueue time, not SM execution time. #1230 also removed kernel-name symbolication from metrics because of cardinality, and pointed this kind of detail at the OTel profiler.

We opened an interest-check on the eBPF profiler (this repo has Discussions; that one does not, so it is an issue):

open-telemetry/opentelemetry-ebpf-profiler#1813

Question for this repo: should OBI stay metrics-only for CUDA, or is there any overlap (shared uprobes, export, discovery) you would want coordinated if the profiler takes this on?