refactor(kernels): migrate kv_cache + audio + vision families

[ekryski]

What

Wave-2 grouped port: three small, clean families migrated into kernels/ in one
PR (you OK'd grouping when it makes sense). Each is its own commit.

• kv_cache — mlx/fft.rs + ffai/{kv_cache,kv_cache_update_many} →
  kernels/kv_cache/ (kv_cache.rs→cache.rs to avoid module_inception).
  mt_ prefixes on the bare kernels (mt_kv_cache_update(_many),
  mt_quantize_kv*, mt_bulk_dequant_kv*); fft already mt_*.
• audio — ffai/{lstm,mel_spectrogram,snake1d,upsample_nearest1d,vocoder} →
  kernels/audio/ with mt_ prefixes.
• vision — ffai/{avg_pool2d_nhwc,frame_diff_luma,im2col_patch,im2col_patch_interleaved, patch_unfold_qwen,pos_emb_2d_add,resize_normalize,transpose_th} →
  kernels/vision/. patch_unfold_qwen → patch_unfold / mt_patch_unfold
  (Qwen stays a usage note in the comment; test/bench fns de-modelled too).

Plus the plan + KERNEL_AUDIT updated to mark these three done.

Generated MSL per kernel is unchanged — only module paths + inventory names.

Verification

• cargo clippy --workspace --all-targets -- -D warnings clean.
• tile test green: kv_cache + quantize_kv + fft; mel/lstm/snake/vocoder/upsample;
  resize/im2col/patch_unfold/transpose/avg_pool/frame_diff/pos_emb. The
  kv_cache_quant_roundtrip integration test passes.

Note on stacking

Stacked on gemm-dense (#20) → ops (#18) → … → dev. (This branch also carries a
fix that #20's force-push picked up: kernels/gemm/gemm.rs → dense.rs to clear
a module_inception that was failing #20's -D warnings clippy.)

[TheTom]

@ekryski clean. kv renames for downstream: kv_cache_update[_many]→mt_kv_cache_update[_many], quantize_kv/bulk_dequant_kv→mt_*. Nit: tests/kv_cache_quant_roundtrip_gpu.rs:6 doc comment still says "ship in ffai::kv_cache". Approve.