Expose cooperative generation and Core ML placement - #461

Summary

• add token-boundary cooperative cancellation to Gemma 4 streaming generation
• let Nemotron callers select Core ML compute units while preserving .all as the default
• add opt-in real-model gates for cancellation and CPU+ANE output parity

Validation

• curated unit suite: 233 passed, 6 skipped
• Gemma cancellation E2E: passed with the cached E4B int4 checkpoint
• Nemotron placement E2E: exact output parity; .all 1465.81 ms, CPU+ANE 1071.30 ms
• release build completed before the worktree transfer

Risk

Defaults are unchanged. Cancellation is cooperative between generated tokens; prompt prefill and each individual token evaluation remain atomic.