Proposal: simdgroup-optimized Metal compute kernels

[cluster2600]

Motivation

The existing Metal kernels in zvec use basic threadgroup reductions with shared memory barriers. Apple's simdgroup cooperative intrinsics (simd_sum, simd_min, simd_shuffle) enable hardware-level reductions across 32 SIMD lanes without shared memory, which should be faster on all Apple GPU families >= Apple4.

Proposed approach

Add 6 new Metal kernels under src/ailego/gpu/metal/:

Kernel	Purpose
metal_l2_distance_simdgroup	32-thread cooperative L2 distance
metal_inner_product_simdgroup	32-thread cooperative dot product
metal_cosine_simdgroup	Cosine similarity with joint norm computation
metal_topk_simdgroup	Parallel top-k selection using simd min-reduce
metal_batch_distance_simdgroup	Fused batch distance for multi-query search
metal_fused_knn_simdgroup	Full fused KNN (distance + top-k in one dispatch)

All kernels would be standalone .metal files with a C++ test. No Python integration in this step — that would come separately.

Questions for maintainers

1. Is this the right directory (src/ailego/gpu/metal/)? I see existing Metal files there.
2. Should these replace the current kernels or live alongside them?
3. Any preference on Metal Shading Language version (I'm targeting MSL 2.4+)?
4. Per your feedback on feat: add GPU backends, quantization, and search optimizations #166, the Metal backend needs "more integration logic" — what integration points would be most useful?

Draft implementation: #172