refactor(sampling): migrate sampling family into kernels/sampling/

[ekryski]

What

Wave-1 kernel-family port: moves the sampling family (the logits→token
pipeline) out of the legacy mlx/ + ffai/ split into kernels/sampling/ per
docs/specs/KERNEL_CONSOLIDATION_PLAN.md, and adds the mt_ prefix to the
previously-unprefixed ffai/ kernels.

Changes

• mlx/{softmax,sort} → kernels/sampling/ (mt_softmax, mt_sort/mt_merge/mt_sort_segmented).
• ffai/sampling.rs → kernels/sampling/categorical_sample.rs; softmax_categorical_sample
  → mt_softmax_categorical_sample (file renamed off the family name to avoid
  module_inception).
• ffai/logits_{topk,top_p,min_p,processors} → kernels/sampling/ with mt_
  prefixes (mt_logits_topk_mask, mt_logits_top_p_mask, mt_logits_min_p_mask,
  mt_logits_temperature, mt_logits_repetition_penalty).
• Excluded from this family: dsv4_*_topk (moe/sdpa expert/indexer routing) and
  upsample_nearest1d (audio) — they go with their own families.
• Updated the categorical-sample integration-test imports, cross-ref doc comments
  (dsv4_indexer_topk, the SPIR-V codegen note), the plan (sampling marked done),
  and the KERNEL_AUDIT rows.

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

Verification

• cargo build -p metaltile-std + cargo clippy -p metaltile-std --tests clean.
• All sampling kernel tests pass via tile test (softmax, sort, the four logits
  masks/processors, categorical sampler).
• softmax_categorical_sample_matrix integration test passes.

Note on stacking

Stacked on the norm PR (#16) → conv+rope (#15) → docs (#14); the diff includes
those until they merge. Each family lands as its own PR and consolidates cleanly.

[TheTom]

Good consolidation: the five logits_* modules folding into kernels/sampling/ with mt_ names matches the plan's one-file-per-op-family-member rule, and the simple elementwise kernels (mt_logits_temperature, mt_logits_repetition_penalty) follow the generic <T> + f32-accumulate house pattern. No issues found in the increment over #16. The KERNEL_AUDIT.md sync in every PR of this stack is appreciated, it keeps the plan document honest as the waves land.

[TheTom]

@ekryski clean. Nice catch repointing the dangling spirv doc cross-ref (crate::mlx::sort::mt_sort → kernels::sampling::sort). One conform heads-up for consumers: ffai_logits_{topk,top_p,min_p} gained a _mask suffix (mt_logits_*_mask). Approve.