[fix][train] Reject unsafe MTP prefix caching

[bvolpato]

Summary

• Reject MTP speculative decoding with prefix caching on recurrent linear-attention models.
• Detect model capability from Hugging Face layer_types instead of model names.
• Respect final engine_init_kwargs overrides for prefix caching and speculative decoding.
• Disable prefix caching in the Qwen3.5 MTP recipe.

Why

Issue #1981 reports a Qwen3.6-27B accuracy collapse when MTP and prefix caching are combined. Current SkyRL config validation accepted that pair because prefix caching defaults to enabled. vLLM issue #43559 tracks the same failure class; proposed fixes #43650 and #50021 remain open and are not part of SkyRL’s pinned vLLM 0.26.0 release.

This guard fails during config validation and gives both supported mitigations: disable prefix caching or disable MTP. Full-attention MTP models and non-MTP hybrid models remain allowed.

Validation

• Pre-fix CPU config reproduction: unsafe Qwen3.6 combination accepted.
• Live Qwen3.6 config: unsafe pair rejected; MTP without prefix caching accepted; SKYRL_DISABLE_SPEC=1 accepted.
• Focused config suite: 163 passed.
• Negative control with guard bypassed: 2 expected regression failures.
• Ruff, Black, and gitleaks pre-commit hooks
• uv lock --check
• Recipe bash -n
• git diff --check

Fixes #1981

[gemini-code-assist]

Code Review

This pull request introduces a validation check to prevent using MTP speculative decoding with prefix caching enabled on models with recurrent linear-attention layers, which is unsafe. It also updates the configuration documentation, disables prefix caching in the example script, and adds corresponding unit tests. The reviewer recommended wrapping the model configuration loading (AutoConfig.from_pretrained) in a try...except block to prevent training from crashing in offline or air-gapped environments if the Hugging Face Hub is unreachable.