fix(server): support sampled requests in target layer split

[weicj]

Summary

From the user side, target layer split currently works for greedy requests but rejects normal sampled OpenAI-compatible requests with sampling_unsupported. In practice this means requests using temperature > 0 such as temperature=0.7, or requests using repetition / frequency / presence penalties, fail even though the same model can sample correctly on the single-card backend.

This PR fixes that gap for target layer split. Adapters that can return final-token logits may opt into the existing CPU sampler path, so split and non-split generation follow the same sampling rules. Adapters that have not implemented logits output still fail explicitly instead of silently falling back to inconsistent behavior.

Changes

• Add a LayerSplitAdapter::supports_cpu_sampling() capability gate.
• Keep the existing sampling_unsupported protection for adapters that do not opt in.
• Extend Qwen35 layer-split projection so it can return final-token logits as well as argmax tokens.
• Extend Gemma4 layer-split projection in the same way, including Gemma4 final logit softcap before CPU sampling.
• Make Qwen35 and Gemma4 layer-split adapters cache the final prefill logits and use the existing shared sample_logits() path during AR decode when temperature > 0, repetition penalty, frequency penalty, or presence penalty is active.
• Persist and restore the cached final prefill logits with layer-split snapshots, so sampled restore paths do not fail or reuse stale logits when no new prefill is needed before decode.
• Keep Qwen35 DFlash speculative decode on the greedy path only; sampled / penalty requests now stay on AR decode where CPU sampling is available.
• Add unit coverage for the layer-split sampling capability gate.

Notes

• This does not add a new sampler. It reuses the existing shared CPU sampler so split and non-split paths follow the same sampling rules.
• Current validation: CUDA SM61 dflash_common, test_server_unit, and dflash_server build passed; test_server_unit passed with 1606 assertions and 0 failures. HIP gfx906 dflash_server build passed, and a dual Pro VII Gemma4 E4B Q4 layer-split smoke returned an OpenAI-compatible completion with temperature=0.7.

[cubic-dev-ai]

2 issues found across 11 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>