Add fused batch distance kernels for vector search

Parent: lance-format/lance#6820
Related: lance-format/lance#6821
Linear: lance-format/lance#6823
Linear parent: lance-format/lance#6820

Goal

Add low-level distance kernels that accept a batch of query vectors and compute distances against a data batch in one fused operation.

Batch query should not only mean shared scan/planning. The hot distance path should avoid a loop that slices one query vector at a time and calls the existing single-query Arrow distance function repeatedly.

Scope

• Add a kernel API in the vector distance / linalg layer that accepts multiple query vectors plus one data vector batch.
• Return distances in a layout that downstream top-k code can consume without excessive reshaping or allocation.
• Cover common dense metrics such as L2, cosine, and dot where applicable.
• Integrate the fused kernel into flat KNN batch query execution.
• Keep scalar fallback paths for unsupported types or platforms.

Acceptance criteria

• Unit tests cover output ordering/layout for multiple queries and multiple rows.
• Flat KNN batch query uses the fused kernel where supported instead of per-query kernel loops.
• Benchmarks compare repeated single-query kernels, current batch query behavior, and fused batch kernels.
• The benchmark report includes dimensions, query batch size, row count, CPU target features, and p50/p95/p99 latency.