engine: hoist prefill sliding-window mask build out of the per-layer attention call

[tlkahn]

Context

Stage D0 phase 2 (#88, PR #91) wires sliding-window attention into fwd_attention. On prefill, the window mask (fwd_sliding_window_mask) is built inside the per-layer attention call, so an identical [1,1,q_len,kv_len] mask is rebuilt for every local layer in the model on every prefill chunk.

Deferred in both review rounds of PR #91: the clean fix needs per-step scratch state owned by model_forward, which does not exist yet.

Proposal

Build the mask once per forward step (keyed on q_len / kv_len / window, which are identical across local layers of a uniform- or pattern-windowed model) and pass it down to the per-layer calls. Natural home is a model_forward scratch struct, which llama3 freqs precomputation (D1) will also want.

Related

• Engine Inference Core Stage D0: config-driven dense-forward foundation + validation/parity scaffolding #77 (D0 foundation, where the seam landed)
• Engine Inference Core Stage D2: mistral - sliding-window attention #79 (mistral D2: every layer is local, so the rebuild cost is num_layers x per-chunk there)
• Engine Inference Core Stage D4: gemma3 - embedding scaling, sandwich norms, local/global interleave, dual rope bases #81 (gemma3 D4: local/global interleave)