fix: Replace std::vector with std::array for distance impl

[iaojnh]

No description provided.

[greptile-apps]

Greptile Summary

This PR replaces std::vector with std::array for all fixed-size local buffers in distance computation hot paths across five files (lloyd_cluster.h, mips_euclidean_distance_matrix.h, and the three inner_product_distance_batch_impl* variants). Since every array size is a compile-time template parameter (BatchCount, M, N, dp_batch), the substitution is valid.

Key observations:

• All converted arrays are properly initialized before use: via std::fill, SIMD setzero intrinsics, or element-wise assignment in loops
• Heap allocations inside SIMD hot loops are correctly eliminated, delivering a measurable performance improvement
• No remaining uses of std::vector or #include <vector> in active code paths
• The commented-out avx512bw fallback (explicitly marked "TODO: this version is problematic") contains issues but is deferred work

This PR is safe to merge; all active code is correct and the change is a net performance win.

Confidence Score: 5/5

• All active code paths correctly initialize and use stack-allocated arrays; this is a safe performance improvement.
• All array substitutions are verified correct: compile-time sizes, proper initialization before use, no uninitialized-read paths in active code. Heap allocations are correctly eliminated from hot SIMD loops. The only code with latent issues is explicitly commented-out and already flagged by the author as deferred work.
• No files require special attention

_{Last reviewed commit: ed3e923}

[iaojnh]

@greptile