feat(sensenova): add vLLM-Omni rollout

[CjhHa1]

Summary

• add a SenseNova-U1.5 vLLM-Omni adapter and RL-aware pixel-flow pipeline with driver-authored noise, FlowGRPO SDE transitions, and replay-ready prefix-cache capture
• add grouped T2I serving, full-weight IPC synchronization, and a colocated training recipe
• bound rollout transfer memory by storing only replay-required trajectory positions and sharing same-prompt prefix caches
• move the vLLM stack to the first stable SenseNova pixel-head release (vllm / vllm-omni 0.22) with an explicit CUDA 12.9 wheel and documented glibc floor

Depends on #389, which adds the SenseNova-U1.5 training-side model and replay implementation. Until #389 merges, this PR intentionally includes those stacked commits; GitHub will drop them from this diff afterward.

Related Issue

N/A

Test Plan

• python3 -m pytest -q tests/models/test_sensenova_u1.py tests/rollout/test_sensenova_vllm_omni.py (13 passed)
• Hydra compose: python3 -m unirl.train_diffusion --config-name=diffusion/sensenova_u1_5/sensenova_u1_5_vllmomni_full_ipc --cfg job --resolve
• pre-commit run --all-files
• H20 end-to-end smoke using SenseNova-U1.5-8B-MoT-Preview, vLLM-Omni 0.22.0+cu129, 1 GPU, 4 denoise steps, one SDE step, and a synthetic two-sample prompt batch: 2 images, sparse trajectory (2, 3, 16, 3072) with indices [1, 2, 4], SDE log-prob (2, 1), shared CPU/pickle-safe prefix caches

Compatibility / Risk

• The vllm extra moves from 0.20 to 0.22 and selects the explicit CUDA 12.9 x86_64 wheel.
• vLLM 0.22's required llguidance wheel raises the effective glibc floor to 2.31; this is documented in INSTALL.md.
• Existing v0.20-only HI3 cache compatibility code now self-skips safely when its removed classes are absent.
• Qwen3-Omni's legacy audio-video M-RoPE correction is gated when native v0.22 Thinker support is detected, avoiding a double position shift.
• SenseNova full-weight sync now fails closed when an input name is neither direct-loadable nor covered by the fused QKV/gate-up mapping.
• The first SenseNova rollout path is T2I, TP=1, full-weight IPC sync; image editing and TP rollout are not exposed by this recipe.

Reviewer Notes

• Review the sigma/data-time sign conversion and unit-noise SDE transition in RLSenseNovaU1Pipeline first.
• Distinct prompts are intentionally split into separate engine calls because upstream SenseNova 0.22 only batches multiple outputs of one prompt.
• Sparse trajectory_indices are carried through the shared vLLM-Omni response builder into LatentSegment.indices for trainer replay.
• This PR is stacked on feat(model): add SenseNova U1.5 pixel-flow support #389.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.