refactor: retire mlx/ — steel attention → sdpa, drop stale steel/conv dupes

[ekryski]

The last legacy folder. Retires mlx/ entirely — lib.rs is now just kernels/ + probe/ + utils/.

What moved

• kernels/sdpa/steel_attn/ (new) — the 4 Steel attention templates (steel_attention{,_mma,_mma_bf16,_nax} = mt_sdpa_prefill{,_mma,_mma_bf16,_nax}), already mt_-prefixed. Intra-refs + the 2 consumer tests repointed mlx::steel::attn → kernels::sdpa::steel_attn.

What was dropped

• mlx/steel/conv/ (steel_conv / _3d / _general) were byte-identical stale duplicates of the already-migrated kernels/convolution/steel_conv/ (doc-only "NOT YET IMPLEMENTED" placeholders — the implicit-GEMM conv needs simdgroup-matmul + im2col DSL primitives that don't exist yet). Deleted.
• The now-empty mlx/ module + its mod.rs/steel/mod.rs.

Verification

clippy --all-targets -D warnings clean; full tile test = 1394 suites / 4125 tests / 0 failed; steel_attn external suites (sdpa_prefill_mma_long_t, _bf16_gpu) pass.

With this, mlx/ and ffai/ are both fully retired — the entire kernel corpus lives under kernels/<family>/. Stacked on #27.

[TheTom]

@ekryski clean, steel→sdpa is a pure relocation (99-100% rename similarity, kernels unchanged), dropped steel/conv are dead dupes, registry-consistency test retargeted to src/kernels. Title nit: "steel attention → sdpa" reads behavioral; it's a folder move, one body line that the templates are preserved verbatim would help.

Cross-cutting CI follow-up (caused by this PR + #27 removing the dirs, not a defect in the move itself): the bench-regression gate goes no-op post-merge, tile.yml shards bench-ffai/bench-mlx gate on src/ffai/**+src/mlx/** paths and --match-group 'ffai'|'mlx', and kernel_family() (harness.rs ~L1364) derives family from the first dir under /src/, which is now always "kernels", so --match-group by family matches nothing and its unit test (~L1463) still asserts mlx/ffai. tile build/tile test aren't family-gated so correctness CI is fine; it's only perf-regression coverage that goes dark. Suggest: have kernel_family() skip a leading kernels/ segment + return the family below it (the path-side counterpart to the op_group fix in #20), and redefine the bench shards/path-filters around kernels/<family>/. Also codecov.yml:78-79 + coverage.yml:59 still ignore src/{ffai,mlx}/**, realign to src/kernels/**. Approve the move; these are a fast-follow.