x86 repacked matmul: add a Q5K tile #4029

On x86, repack_x86.rs::select() repacks only Q4K, Q6K and Q8_0, so Q5K weights take the generic k_quants::matmul path. aarch64 already packs Q5K (pack_to_q5kx8). In my #4003 measurements, Q5K prefill (2048x2048, m=128) ran at ~11.6k rows/s against ~31k rows/s for repacked Q4K on the same machine, and Q5_K_M models store most weights as Q5K.

Proposal: a TileQ5K mirroring TileQ4K (16 rows, unpacked scales/mins, nibbles plus the high bit merged into u8 0..31), reusing the existing dpbusd/maddubs loop with the same min correction, gated to m > 1 like Q6K so decode keeps vec_dot_q5k_q8k. Would you take a PR for this? @EricLBuehler