fix(attention): correct causal mask for multi-tile decoder attention (seq_len > mlen)

[booth-algo]

The native per-head flash-attention kernel (_flash_attention_mha) added the same static (mlen, mlen) triangular causal mask to every QK score tile. That is correct only when seq_len <= mlen, where the seq x seq score matrix is a single tile. When seq_len > mlen the score matrix tiles into a grid and the mask must depend on tile position: strictly-future key tiles (k_idx > q_idx) must be fully masked but the triangular mask left their lower triangle open (leaking future keys), and strictly-past key tiles (k_idx < q_idx) need no mask but the triangular mask wrongly masked their upper triangle (dropping valid past keys). Only the diagonal tile was correct, producing a partial numerical divergence (~58% allclose) on decoder runs at seq_len > mlen.

Fix

The kernel now derives each tile's global causal geometry: strictly-future key tiles contribute exp(-inf)=0 and are skipped entirely, the triangular mask is applied only on the diagonal tile, and strictly-past tiles are left unmasked. A NotImplementedError guards the unsupported kv_seq_len != seq_len (KV-cache query offset) case rather than silently miscomputing. seq_len <= mlen remains the single-tile special case and is byte-identical; non-causal callers (causal_mask=None, e.g. the vision encoder) are unaffected.

Why it was never caught

This regime was simply never exercised. The sub-64 decoder fixes (#54/#55) validated seq_len=4 < mlen (single tile); the default mlen=256 keeps seq_len=64 < mlen (single tile); vision is non-causal. The first run at seq_len > mlen for a causal decoder was the recent sub-64 blen sweep, which surfaced the divergence.

Verification

Isolation run, identical config native_64x64x16, only seq differs — seq=64 (1 tile) PASS vs seq=128 (2 tiles) FAIL→PASS — pins the cause to multi-tile causal masking, ruling out the #56 RoPE/norm rolling (same rolled code passes at 1 tile). The full SmolVLM2 1-layer decoder sweep at mlen=16 and mlen=32 across all blen goes from allclose FAIL to PASS:

mlen	blen	sim_lat	allclose
16	4	19.74ms	PASS (97.8%)
16	8	6.86ms	PASS
16	16	3.63ms	PASS
32	4	16.72ms	PASS (98.5%)
32	8	4.76ms	PASS
32	16	1.76ms	PASS
32	32	1.01ms	PASS

The existing flash_attention_mha golden test still passes at 100%, and a new codegen regression test (test_mha_causal_skips_future_tiles_and_masks_only_diagonal) asserts that strictly-future tiles are skipped and the triangular mask lands on diagonal tiles only — it fails on the pre-fix kernel.