perf(rocprofiler-sdk): kernel replay phase 2: performance optimizations

[MythreyaK]

Motivation

Experimental performance improvements for kernel replay. For now, this branch focuses on exclusive device-side snap/restore (i.e., is device memory is full, replay fails).

Upcoming commits might include graceful host-side fallback etc.

Technical Details

Performance optimizations on top of #7960 kernel replay. As implemented,

• snap on device memory (only, for now)
• implement blit copies batched with kernel dispatch for gap-free restores, without host-side API call for restoring memory
• signal reuse for dispatches (saves ~200us)

Overall, on a simple babel stream benchmark, runtime improves ~9.4x. Concrete results and a formal rocprofiler-sdk/benchmark suite will be added after implementation stabilizes.

Issue Tracking

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.