perf(qwen3): skip redundant decode d2h/host copies in greedy path

[lyfne123]

Summary

Stacked on hw-native-sys#68 (needs its decode_fwd restore + pypto-lib bump). The fused Qwen3-14B decode kernel does device-side embedding and greedy sampling, so several per-step host copies are pure waste. This removes three, all host-only — no kernel change, no recompile. All Qwen entry points default to temperature=0 (greedy device sampling), so this is the hot path in normal runs.

#	Copy eliminated	Per decode step	When
①	greedy logits d2h	16 × 152064 × 4B ≈ 9.7 MB	greedy
②	next_hidden d2h	16 × 5120 × 2B ≈ 160 KB	always
③	dead decode hidden copy	[16, 5120] alloc + copy + float	always

How

The runtime only device→host copies host shared-memory args; worker-resident DeviceTensor args stay on device. So:

• ① logits — in the device-greedy path the token id comes from decode_sampled_ids_buffer (device), and host code never reads the full [BATCH, VOCAB] logits. Point the kernel's logits out slot at a reusable worker-resident DeviceTensor (_decode_device_scratch) instead of the shared host buffer. The kernel's internal argmax stays device-local; the d2h disappears. Non-greedy still uses the host buffer and returns real logits.
• ② next_hidden — decode never returns it (run_decode always passes None to _integrated_sample_result), so its Out buffer is scratch. Route it to a device-resident buffer unconditionally.
• ③ dead hidden — _prepare_decode_inputs built a [B, hidden] bf16 buffer from batch.hidden_states (a zeros placeholder for device-embedding executors) that the kernel does not consume and no caller reads. Stop building/returning it.

DecodeResult.hidden_states / .logits become Optional; serving_worker reads host logits lazily (only on the non-greedy fallback), so None is safe. engine._sample_batch_rows already short-circuits on device-sampled ids before touching logits.

Verification (NPU, a2a3, greedy temperature=0)

• Correct output: "The capital of France is" → "Paris … Washington", 24 tokens over 23 decode steps through the new path, no errors.
• Run-to-run token divergence was confirmed to be inherent NPU non-determinism: the same unmodified binary run on two devices diverges identically at a near-tie argmax (step ~10). Not caused by this change — logits argmax reads the same values whether the buffer is host or device, and next_hidden is write-only.

Scope

• Non-greedy / streaming paths unchanged (still return host logits).
• Prefill's full-embedding h2d is not touched here — it needs kernel-side device gather.