Vulkan Optimizations and Fixes

[0cc4m]

I have implemented a number of Vulkan optimizations and fixes:

• Implement REPEAT operator shader to fix low performance of Vulkan copy-based implementation
• Use GLSL FMA instruction where possible
• Add GGML_VULKAN_PERF option to get approximate performance data about a running model
• Rework and fix Vulkan Descriptor Set handling, this improves performance in my tests on AMD RADV
• Fix validation error on float32 concat f16 shader

I will keep this on draft while I check a few more things, but feel free to test and benchmark. Don't expect a huge difference.

---

• [x] I have read the contributing guidelines
• Self-reported review complexity:
  • [ ] Low
  • [x] Medium
  • [ ] High

[0cc4m]

I missed a validation issue in #8943, but the fix is now in this branch. I think this should be ready for a review and then merge.

[0cc4m]

@ggerganov @slaren Can one of you review the non-Vulkan parts of this PR and approve if that's fine?