[bitnet] Wire BitNetTwoStageDecode into the generation loop (pre-lm_head hidden-state hook)

[michalharakal]

Problem

BitNetTwoStageDecode (stage-1 planes-0–3 scan + exact candidate rescoring with a provable top-k-equality bound) and the sampling API sampleFromCandidates/ScoredToken (llm-core/.../SamplingUtils.kt:154-190) shipped with PR #353 — but no generation loop calls them. OptimizedLLMRuntime exposes no pre-lm_head hidden state, and the CLI's BitNet branch decodes single-stage full-vocab. The two-stage path is reachable only from tests and library callers.

This is the residual #337 flagged as "a separate, small design": the hook was never added.

Proposed change

• Runtime: expose the last-position hidden state before the lm_head projection (hook or return-shape addition in OptimizedLLMRuntime), so a family can substitute candidate scoring for the full-vocab matmul. Keep it format-driven — the hook is generic; only the BitNet path uses it today.
• CLI: on the BitNet branch, use two-stage decode when the head weight is BITNET_PLANES (with [bitnet] Tied lm_head never engages the planes kernel on 2B4T — FP32-widened head, ~0.45 tok/s #357 landed this covers 2B4T's tied head). Single-stage remains the default everywhere else.

Verification

• DIRECT and OPTIMIZED produce identical sampled tokens (both dispatch by weight format — [bitnet] T2: lm_head via BITNET_PLANES weight format (WeightForm) + two-stage top-k decode #337's decode-mode coverage requirement).
• Greedy decode: two-stage output identical to single-stage full-vocab on a real BitNet GGUF; tok/s delta recorded.

Refs: #335, #337 (residual), #357 (tied head), SKaiNET-developers/SKaiNET#1150.