Add synchronous fixed-budget Ouro rollouts for RL

[0z5a]

This adds a synchronous RL interface to the existing Ouro engine: one fixed depth for prefill and decode, selected-token full-vocabulary temperature-scaled logprobs, and full-weight publication while the engine is idle. It preserves the native model, scheduler, sampler and per-depth KV implementation. Publication advances the policy version only after copies finish; failed publication leaves generation disabled. Prefix reuse and asynchronous/speculative serving are excluded from this interface.

Base: ThinkFlowLab/vllm-rlt@ea680f7735eeaa357d2e09ba24355747496041dd. This fork draft is a companion to a VIME shared-parameter Megatron provider and global-budget GRPO recipe. It adds an in-process interface, not HTTP publication endpoints.

Validation on 2×H20:

Check	Result
Existing engine/Ouro tests plus RL tests	55 passed
Tiny independent HF gradients, recompute, cached logprobs, publication	Passed at K=2/3/4
Full pinned Ouro-1.4B FP32 cached selected-token logprobs vs dense training forward, K=4	Maximum absolute error 6.68e-6
Full BF16 cached selected-token logprobs vs SDPA training forward, K=2/3/4	Maximum absolute error 0.0478 / 0.1057 / 0.1046
Actual prefill block-tokens for an 8-token prompt, K=2/3/4	384 / 576 / 768
Actual decode block-tokens for seven cached steps, K=2/3/4	336 / 504 / 672

The BF16 paths are not bit-exact. Changing the HF causal-mask dispatch alone also changes BF16 logits; the matched SDPA training/HF paths are bit-exact and FP32 default-HF/dense training logits match. Real RL quality and time-to-reward improvement are not established by these numerical tests. The draft remains experimental pending the real short reward pilot.