feat: implement Gemma2 sliding-window attention (local/global alternation)

[inureyes]

Part of #493

Context

Gemma2 alternates sliding-window attention (window 4096) and global attention across layers. The emitter currently implements global attention only, which is a no-op difference for prompts shorter than the window but produces wrong results once the context exceeds 4096 tokens.

Scope

• Implement the per-layer sliding-window mask so the local/global alternation matches Gemma2.
• Wire the window size and the per-layer local/global selection from Config::from_json.
• Files: emitter/model.rs, emitter/config.rs.

Acceptance criteria

• Token-exact vs the HF fp32 oracle on a prompt longer than the window. Validated at reduced scale: the emitter caps context at MAX_SEQ = 256 (< the 4096 production window), so the check runs the identical mask code with a shrunk window that bites inside 256. A synthetic Gemma2 (matched HF config) is token-exact and bit-exact (max |logit diff| ~6e-8) vs the HF fp32 oracle at both an inert window (>= prompt) and a biting window (< prompt), and HF's own output changes between the two, so the biting case is not vacuous. The literal >4096-token GPU sweep needs MAX_SEQ > 4096 (out of scope here) and is the orchestrator's post-merge gate.
• Existing short-prompt Gemma2 outputs unchanged (token-exact single-seq, reference-exact serve). For W >= MAX_SEQ the window predicate is always true, so the local mask equals the causal mask on the value; the inert-window oracle run is bit-exact vs HF, and the Llama-3.2-1B byte-for-byte asset test is unchanged.
• Integrated: the sliding-window mask is emitted in all three graph kinds the CLI (MLXCEL_BACKEND=xla) and mlxcel-server compile from config.json at load (single decode, prefill, ragged decode), so a Gemma2 checkpoint runs with the local/global alternation on both paths. Long-context behaviour beyond 256 tokens is bounded by MAX_SEQ (a separate follow-up), not by the masking.

Validation

• Pure-Rust structural tests (scoped cargo test -p mlxcel-xla --lib emitter::tests): window parse + default, is_sliding_layer even-local schedule, one-time local-mask block per graph, window value emitted, and per-layer alternation (even layers consume the local mask, odd the global, and they differ). Llama byte-for-byte asset test still passes (non-Gemma2 emission byte-identical).
• Independent HF-oracle execution check: spike/openxla/gemma2_sliding_window_check.py (emit -> IREE llvm-cpu -> compare to HF eager fp32).

Dependencies

None required. Benefits from #494 (shared attention core) if it lands first, but can proceed independently.