Improve Gemma4 long-context correctness and Metal decode performance

[timkaye11]

[codex]

Summary

This PR improves Gemma4 E2B/E4B long-context inference

• Fixes mixed sliding-window/global-attention KV behavior, bounded chunked prefill, paging, cache growth, and ownership for long prompts.
• Replaces the fixed 2,048-token request cap with model-aware limits and route-aware KV admission that reserve output capacity and fail conservatively outside verified paged Metal paths.
• Applies Gemma4 partial-rotary RoPE theta exactly once across native and Metal execution while preserving long-context MTP and pipelined-frame correctness.
• Adds a bounded, fail-safe Metal split-GQA decode path for supported Gemma4 shapes, with telemetry, a kill switch, and automatic fallback.
• Hardens shared generation admission with bounded speculative width, exact Qwen3.5 image-token accounting, expanded-context revalidation, and consistent benchmark ignore_eos handling.
• Keeps short-context, unsupported shapes, and non-paged fallbacks on their existing conservative paths.

Benchmarks

Local, unattested Apple Metal measurements use ReleaseFast builds and fresh processes. The comparator is the branch's exact main fork point; decode excludes prefill.

Workload	Main	This PR	Result
Gemma4 E2B, 470+64	34.50 tok/s	39.98 tok/s	+15.9% decode; 6.7% lower total time
Gemma4 E2B, 2,010+64	34.78 tok/s	49.84 tok/s	+43.3% decode; 4.1% lower total time
Gemma4 E4B, 444+64	21.26 tok/s	21.43 tok/s	Performance parity; 0.8% lower total time
Gemma4 E4B, 2,002+300	18.85 tok/s	26.93 tok/s	+42.9% decode; 21.1% lower total time

An additional E2B 2,065+300 long-output check used one warmup and five measured fresh processes:

TTFT	Prefill	Decode	Generation	Total
7,946 ms	259.88 tok/s	49.64 tok/s	13,990 ms	14,647 ms

Validation

• Short E2B/E4B controls produced identical token IDs versus main.
• Long E2B and E4B outputs were token-exact with the optimized Metal route enabled or disabled.
• E4B 2K+300 recorded all 12,600 expected split-GQA dispatches with zero fallbacks.
• Focused admission, bounded-prefill, generation-route, and memory-estimator tests passed 7/7.
• Default Metal, -Dmetal=false, and ReleaseFast Metal builds passed.
• The E2B 2,065+300 run completed deterministically at the requested 300-token length.
• OpenAPI assembly is current, including the supported speculative-width bounds.