Fix ANE-LM review findings after PR #86

[IchenDEV]

Summary

• disable Qwen3 thinking in the direct ANE prompt so formatting and structured edit-command requests spend their token budget on the requested output
• reject ANE requests that exceed the packaged runtime's 2,048-slot KV cache, reserving the slots needed by generated tokens so oversized requests enter the existing MLX fallback path instead of silently dropping the prompt prefix
• unload the ANE runtime before MLX fallback for both generation and warmup paths
• preserve the first fallback outcome in the environment-gated repeated-request test

Regression coverage

• exact non-thinking Qwen3 assistant suffix
• 2,048-slot boundary and output-headroom cases
• ANE cleanup occurs before MLX starts, while disabled fallback performs no cleanup
• real fallback test asserts the ANE engine is unloaded and captures .fallback before later MLX iterations clear task-local state
• environment-gated real ANE lifecycle test rejects thinking-tag output and requires the first structured command response to parse

Follow-up to #86 and its unresolved review threads.