refactor(norm): migrate normalization family into kernels/norm/

[ekryski]

What

Wave-1 kernel-family port: moves the normalization family out of the legacy
mlx/ + ffai/ split into kernels/norm/ per
docs/specs/KERNEL_CONSOLIDATION_PLAN.md, and drops the ffai_ / bare-name
prefixes so kernels are named for the operation, not the source.

Changes

• mlx/{rms_norm,layer_norm} → kernels/norm/ (mt_rms_norm*, mt_layer_norm).
• ffai/{rms_norm_residual,rms_norm_rope,gated_rmsnorm,rms_norm_qgemv,gated_rms_norm_qgemv}
  → kernels/norm/ with the ffai_ prefix dropped (mt_*).
• ffai/adain1d: adain1d → mt_adain1d (it had no mt_ prefix at all).
• The two *_block_scaled_qgemv format matrices move as-is — they already use
  clean mt_<format>_* names; the format-axis fold (plan §7) is a later pass.
• Updated the two integration-test imports, the batched_qkv_qgemv /
  rms_norm_block_scaled_qgemv cross-ref doc comments, the consolidation plan
  (norm marked done), and the KERNEL_AUDIT norm rows.

Generated MSL per kernel is unchanged — only module paths and inventory names
change. The FFAI emit consumer is regenerated from the new inventory.

Verification

• cargo build -p metaltile-std + cargo clippy -p metaltile-std --tests clean.
• All norm-family kernel tests pass via tile test (rms_norm, layer_norm,
  adain1d, gated, residual, rope, and the block-scaled format matrices).
• rms_norm_per_head_gpu and dispatch_pinning_guard integration tests pass.

Note on stacking

This branch is based on the conv+rope reorg branch (#15), which in turn is based
on the docs branch (#14), so the diff here includes those until they merge. Each
family lands as its own PR; they consolidate cleanly once the earlier ones merge.

[TheTom]

Increment over #15 is clean (~100 lines, mostly moves + doc sync). The fused mt_mxfp4/nvfp4/mxfp8_e4m3_gated_rms_norm_qgemv family keeps one kernel per format for now, which is the right interim call: folding the format onto a variants(FMT = ...) axis is wave-3 quant work per the consolidation plan, and doing it here would bloat a move-PR with body rewrites.

One small thing: mt_adain1d ships with f32 tol 1e-3 where the house default for f32 is 1e-4. If the looser tolerance is needed (long reduction or transcendental in the body), a one-line comment on the #[test_kernel] saying why prevents the next person from tightening it and chasing a phantom failure.

[TheTom]

@ekryski clean, all 13 norm files moved + re-exported, block-scaled qgemv correctly left mt_-named, non-move edits are doc/import only. Approve.