inference: Qwen3.5 support (VL unwrap, Gated-DeltaNet LoRA targets, linear_attn walk)

Phase A serving speedup: zero-copy state-cache expand (no deepcopy per decision)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)

inference/README.md: mention typical-medium alongside small/preview

Add inference/: minimal self-contained Typical inference package