perf(vllm-omni): batch BAGEL grouped t2i rollout into one packed generate_image

[zzhuoxin1508]

Summary

• Group BAGEL vLLM-Omni t2i rollout requests by prompt and use
  num_outputs_per_prompt=samples_per_prompt (spp) instead of sending one diffusion
  request per image sample. The shared prompt KV is built once and replicated ×spp;
  all spp images are denoised in one packed generate_image per step, then the
  engine splits them back per request.
• Keep downstream UniRL semantics sample-level by preserving sample-level x_T
  recipes (grouped-span slice of resolve_request_noise), splitting log-prob +
  trajectory per image ([spp, T+1, seq, C]), and validating the grouped prompt
  layout via grouped_texts_from_req.
• **Bugfix ** feat(trainer): add periodic eval to PE / ReFL / UnifiedModel #202 renamed the
  pack_initial_noise_extra_args() keyword n_prompts= -> n_samples= and updated the
  shared dit.py caller, but BagelInputAdapter (merged from a side branch) still
  passes n_prompts=. Since that rename, every bagel_t2i rollout that packs a
  driver-authored x_T crashes on its first generate with `TypeError: pack_initial_noise_extra_args() got an
  unexpected keyword argument 'n_prompts'

Compatibility / Risk

• Packed path is scoped to pure bagel_t2i + cfg≤1 (_is_packable_t2i /
  _is_batchable_t2i). cfg>1 / text-output / i2i keep the sample-level layout
  (num_outputs_per_prompt=1); a mismatched num_outputs>1 on the worker raises
  instead of a broken silent fallback. spp=1 is byte-identical to before.
• Missing CFG keys are treated as non-batchable (upstream Bagel defaults absent
  keys to CFG-ON @4.0/1.5).
• Packed images are isolated by varlen block-diagonal attention, so a packed forward
  == spp separate forwards. Only per-step SDE RNG differs (<0.1%, no collapse, ratio 1.0).

Test Plan

• 8×H20 BAGEL vLLM-Omni E2E, 3-way A/B (spp=16):

  variant	e2e s/rollout	ratio
  trainside	~187	1.0000
  unmodified vllm_omni	~150	1.0000
  batched (this PR)	~139	1.0000

• Confirm bagel_t2i with driver x_T (recipe / initial_noise_batch) no longer raises
  on pack_initial_noise_extra_args.

• Smoke cfg>1 + spp>1: stays on sample-level path (no pack).

Reviewer Notes

• Reuses perf(vllm-omni): group SD3 and Qwen-Image prompt rollouts  #117's grouped-span resolve_request_noise — no new noise code.
• Adapter/pipeline pack gates must stay aligned (_is_packable_t2i ↔ _is_batchable_t2i).

[CjhHa1]

LGTM