Decode throughput: move lm_head to ANE for all LlamaPrefill models

[axiom-of-choice]

#215 / #224 folded GPT-2 into LlamaPrefill with host-side lm_head and wpe. speculative.py already has _build_lm_head that runs the head on the ANE as a tiled compile_multi, amortising cost over Kq tokens. That pattern should be generalised to LlamaPrefill.generate so every model (Llama/Qwen/GPT-2/Mistral) benefits, not just the speculative verifier.
Scope:

• Add _build_decode_lm_head to LlamaPrefill, using limit("max_tensor_dim", detect_family()) so A16+ models with vocab < 65536 get a single dispatch.
• Replace _logits host-matmul in generate with ANE lm_head when the model opts in (or by default if numerics stay exact).
• Ensure release() frees the lm_head program.
  Blocked on GPT-2 decode throughput: family-aware lm_head tiling + fold wpe into the graph #226 completing first, so we have the family-aware tile size and a measured GPT-2 baseline to compare against.

[axiom-of-choice]

Will take this after i finish with #226