feat(phi3-small): implement blocksparse attention

[inureyes]

Summary

Replace the dense-attention fallback on Phi-3-small's non-dense layers with the real blocksparse pattern (local block window + per-head vertical stride), while keeping the periodic fully-dense layers on the plain causal path. This makes long-context output follow the trained sparsity instead of diverging toward dense.

What changed

• src/models/phi3small.rs
  • Added build_blocksparse_mask(...): builds the additive [1, n_heads, q_len, kv_len] mask on-device (broadcasting ops only, nothing of size q_len * kv_len materialized on the host). Faithful port of mlx_lm/models/phi3small.py (Attention._block_sparse_mask + _block_sparse_attention), including the reference's right-aligned query-block assignment and the per-head vertical stride (kb + head + 1) % vert_stride == 0. Token-level causality is folded into the same mask so it reproduces the reference mask + dense_mask combination exactly.
  • Split the attention dispatch into dense_attention (unchanged causal path, used by the periodic dense layers) and blocksparse_attention (non-dense layers), threading the already-computed per-layer block_sparse flag into the attention path.
  • In-window fast path: when the whole key length fits inside the local window (kv_len <= num_local_blocks * block_size) the pattern is exactly causal, so the dense path is reused. This gives byte-identical short-context output and, for in-window L == 1 decode, allocates no per-step mask (the maskless single-query fast path selects the whole cache).
  • Stored blocksparse_block_size / _num_local_blocks / _vert_stride on Attention.
• src/models/phi3small_tests.rs (new): config parsing/defaults, block_sparse flag pattern, device-mask vs host-reference agreement (bulk + explicit key-block set), short-context mask==causal and RMS parity against the dense fallback, long-context sparser-than-causal + runtime divergence, and a decode single-row mask check.

Decode (L == 1) note

The issue asks to prefer gather-based KV selection over a full per-step mask. In-window decode does exactly that (no mask; trailing-cache selection via the maskless fast path). Beyond the window, the union of attended blocks across all heads spans the full causal range for Phi-3-small's head count (vert_stride divides the head set), so a per-head gather cannot shrink the fused-SDPA work and would require a ragged per-head kernel; we build a single [1, n_heads, 1, kv_len] additive mask (one row) and let the fused kernel consume it.

Test plan

• cargo test --release phi3small --features metal,accelerate (mask pattern, short-context dense parity, long-context sparsity, decode row)
• cargo clippy --lib --tests --features metal,accelerate -- -D warnings
• cargo fmt --all -- --check

Validation scope (honest)

Validated: unit tests above, including short-context RMS parity with the dense fallback and mask-pattern agreement with the mlx-lm reference formula on constructed cases, run on Apple Silicon Metal. NOT run: generation on a real Phi-3-small checkpoint (no Phi-3-small checkpoint available in this environment; Phi-3-small is a ~7B model and a multi-GB download was out of scope here). Long-context behavior is validated at the mask/selection and SDPA-output level, not via a real-model long-context generation.

Closes #511

[inureyes]

Duplicate of #581, which already squash-merged this exact change (commit 92396a8, phi3-small blocksparse attention) into main. #582 was opened automatically on the same branch after #581 merged and carries no content difference. Closing as superseded.