Chatterbox/Chatterbox-Turbo PR 2: min_p as a per-request sampler knob

[merceod]

What does this PR do?

Adds min-p sampling (HF MinPLogitsWarper, vLLM min_p) as a per-request knob, applied where HF applies it, after the repetition penalty and temperature and before top-k/top-p. Chatterbox's reference sampler runs min_p=0.05 there, and a model can't reproduce that outside the sampler because the penalty is applied inside it.

• SamplingReqConfig.min_p / SamplingConfig.min_p, default 0 (off).
• SamplerSpec.enable_min_p, default False, same idea as enable_repetion_penalty. Only nodes that opt in carry the filter in their captured sampler (a [B] buffer in SamplerBuffers, two extra passes over [B, V] per step). Every other node's graphs and eager path are untouched.
• apply_min_p(probs, min_p) drops probs < min_p * max and renormalises. No CPU branches or data-dependent shapes, so it captures fine. min_p == 0 rows and greedy one-hot rows pass through. It sits after fused_temperature_softmax and before FlashInfer's top-k/top-p. The XPU path masks the raw logits instead, since its kernel samples from logits.
• SamplerResource.ingest_request refuses min_p > 0 on a node that didn't opt in. Otherwise eager steps would honour it and graph replays wouldn't.

How was it tested?

test/modular/test_sampler_min_p.py checks the filter against the HF warper on random logits, identity at 0 and argmax at 1, greedy pass-through, the config reaching the eager sampler, buffers allocated only when enabled, the resource refusal and the spec default. One CUDA-only test (min_p=1 is greedy for every seed on both the eager and graph paths) passed on an H100. docs/adding_models.rst documents the knob. ruff check . and python -m pytest test/modular pass on the rebased branch.

Checklist

• ruff check . passes
• Added or updated tests / docs where relevant