Add small-dimension batch kernels for Hamming and binary metrics

Parent: #6820
Related: #6823
Linear: https://linear.app/lancedb/issue/OSS-926/add-small-dimension-batch-kernels-for-hamming-and-binary-metrics
Linear parent: https://linear.app/lancedb/issue/OSS-922/vector-search-performance-optimization-roadmap

Goal

Add metric-specific batch distance kernels for small-dimensional workloads, starting with Hamming / binary vectors.

For small dimensions, a general batch kernel may not provide enough work per query vector to use SIMD efficiently. Hamming-like metrics need targeted kernels and benchmarks across realistic byte/bit lengths.

Scope

• Identify the existing Hamming / binary distance hot path and benchmark dimensions.
• Add batch kernels for small fixed byte/bit widths where specialization is likely to help.
• Keep portable fallbacks and only use platform-specific SIMD behind the existing target-feature patterns.
• Avoid broad specialization that regresses non-hot dimensions.
• Integrate the specialized batch kernel into batch vector search where the metric/type match.

Acceptance criteria

• Benchmarks sweep representative small dimensions instead of reporting only one favorable size.
• Results separate hot lengths that improve from lengths that are neutral or regress.
• Tests validate distance equality with the scalar/general implementation.
• Platform-specific implementations have fallback coverage for unsupported CPUs.