Fix MTP speculative decoding on multi-GPU splits (fixes #260)

[dushyant30suthar]

Both fixes from #260, as offered there.

1. Group batched rewind jobs by each layer's actual device (modules/gated_delta_net.py)

_collect_rewind_jobs assumed all of a cache's GDN layers live on one device and launched every rewind job under the first layer's index. With layer-split loading the layers span every device in the split, so batched_conv_rewind/batched_state_rewind dereferenced foreign pointers — CUDA illegal memory access on the first draft rejection. Jobs are now grouped per device and dispatched under the right guard/stream. The TP path is unaffected: within a worker all layers share a device, so the grouping degenerates to the previous single-entry behavior.

2. Completion wins over requeue (generator/generator.py)

In the drafting branch of iterate_gen the requeue check ran before the EOS check. A job producing its final token while a requeue condition held was requeued instead of completed; the replacement carried a zero-or-negative remaining token budget, could never reach EOS, requeued itself indefinitely, and each orphan held a recurrent state slot — with the default pool of 4, the fifth distinct request failed on Cannot create new state: no available slots and the generator wedged. rq and not eos restores the intended precedence; non-EOS requeues (output chunking, rollback recovery on long generations) are untouched.

Validation: 2× RTX 5060 Ti (sm_120), native Linux, CUDA 12.8 and 13.3, Qwen3.6-27B EXL3 5.0bpw with embedded MTP tensors, via TabbyAPI and via the bare generator API. 30/30 requests across the failure patterns from #260 (max_tokens=1 series, distinct prompts, growing multi-turn conversation), plus full 4k–163k depth sweeps with drafting active on both layer split and tensor_parallel — no illegal memory access, no slot exhaustion, draft acceptance in the expected range (~2.5–3.1× accept-speedup on code/agentic prompts).

[dushyant30suthar]

Both fixes have landed in dev (ce77f0a and 1048003) — thanks for taking them, and the receive_sample() placement for the requeue/EOS one is the better call since it covers the non-draft path too.

Closing this in favour of those commits. I've been running dev equivalents on the 2× RTX 5060 Ti setup with tensor_parallel + MTP for a full day of real agentic traffic since: 787 requests, ~181k tokens generated, 97.6% prefix cache hit rate, no slot exhaustion and no rewind crashes.