[Serve] Disabling the overlap scheduler changes generation at temperature 0

Description

The same prompts produce materially different generations at temperature=0 depending on scheduler mode. On a 46-prompt classification benchmark (Qwen/Qwen3.5-4B, A10G, identical prompts and sampling params):

Reproducible back and forth: restoring the default scheduler restores 85%. The output distribution shifts noticeably (in our case the model stopped abstaining on low-signal inputs: 'unsure' recall dropped from 10/12 to 6/12). The large latency difference for identical workloads also looks suspicious.

At temperature 0 both modes should be near-deterministic and agree; a per-mode numerics or sampling divergence this large makes results measured in one mode invalid in the other — and since logprobs currently require the non-overlap mode, any logprob-based evaluation inherits the divergence.

Repro sketch: serve Qwen/Qwen3.5-4B --max-length 4096 --max-batch-size 1 with and without the flags above; send identical chat requests with temperature: 0 and compare outputs.

Assisted-by: AI

Reproduce the issue with Qwen/Qwen3.5-4B using the Serve command and the two scheduler flag sets, then compare temperature-0 outputs and latency on the stated prompts. Check the overlap scheduler and non-overlap path, including the logprob-required mode. Done means the scheduler modes no longer produce materially different generations for identical inputs, or the divergence is clearly characterized.