Instrument Vulkan frame profiling

[anirul]

Summary

• add an opt-in Vulkan frame profiler enabled with FRAME_VULKAN_PROFILE=1
• aggregate frame timing and counters every 120 rendered frames
• instrument Vulkan raytrace update, GPU skinning, BLAS/TLAS update, command recording, submit, present, and fence wait paths

Findings from grpcMMO run

Captured with grpcmmo_client.exe --device=vulkan --auto_move_seconds=30 at 1280x720.

Steady average after the first profile window:

• FPS: ~102.8
• renderer.display: ~8.07 ms/frame
• renderer.update_raytrace_buffers: ~7.27 ms/frame
• raytrace.gpu_skinning_update: ~3.20 ms/frame
• raytrace.hardware_transform_update: ~1.74 ms/frame
• raytrace.tlas_update: ~0.84 ms/frame
• raytrace.gpu_skinning_submit_wait: ~0.65 ms/frame
• raytrace.tlas_update_submit_wait: ~0.36 ms/frame

Counters showed raytrace.gpu_skinning_blas_updates=1.00/frame and raytrace.tlas_updates=1.00/frame, so the current slow path is per-frame raytrace buffer/acceleration-structure update rather than present, command recording, or fence wait.

Verification

• cmake --build --preset windows-client-debug
• local Vulkan client profiling run with temporary local auth/game server ports