[Vulkan] Implement vendor/architecture-specific matmul tuning

[goniz]

Problem

MLX uses one default MLX_VULKAN_MATMUL_SPEC tuple unless manually overridden by env var. ggml has a much richer architecture/tuning matrix including vendor detection, subgroup characteristics, cooperative matrix probing, and many precompiled matmul/FA variants.

Tasks

• Build vendor/architecture tuning table with:
  • matmul tile sizes
  • subgroup sizes
  • aligned vs unaligned variants
  • f16acc vs f32acc
  • split-K thresholds
  • small/medium/large kernel families
• Add subgroup-size control and cooperative matrix capability detection
• Add integer-dot-product support detection
• Add architecture-specific kernel selection

Related

See Tier 2 item 1 in performance analysis report.