[Vulkan] Add VkPipelineCache persistence and kernel warmup

[goniz]

Problem

MLX caches pipelines in-process in KernelManager, but there is no VkPipelineCache object, no persistence, and no warmup path. ggml prebuilds many more variants, reducing "compile when first hit" stalls.

Tasks

• Add VkPipelineCache object creation
• Persist pipeline cache blobs to disk per device/driver build
• Prebuild known hot kernels at init or on first model load:
  • matmul, softmax, rope, RMSNorm, FA main paths, quant/dequant

Related

See Tier 2 items in performance analysis report.