[Feature]: Support request-level batching for Wan2.2 pipelines

🚀 The feature, motivation and pitch

Currently, Wan2.2 pipelines (Wan22Pipeline, Wan22I2VPipeline, Wan22VACEPipeline, Wan22S2VPipeline) do not declare support for request-level batching (supports_request_batch = False). When running offline or online inference with multiple requests in a batch, calling execute_model_batch raises:

RuntimeError: Wan22Pipeline does not support request-batch forward.

Supporting request-level batching (supports_request_batch = True) for Wan2.2 will allow batching multiple prompts/requests in offline batch processing and increase GPU utilization when processing concurrent requests of the same resolution and step count (similar to QwenImagePipeline, SD3Pipeline, FLUXPipeline, and LTX23Pipeline).

Proposed changes:

1. Declare supports_request_batch = True in Wan2.2 pipeline classes (Wan22Pipeline, Wan22I2VPipeline, etc.).
2. Update forward(self, req: DiffusionRequestBatch) to accept len(req.prompts) >= 1, removing the single-prompt validation check.
3. Update encode_prompt() and prepare_latents() to support batched prompt lists and stacked latents ($B > 1$).
4. Pass batched inputs through predict_noise_maybe_with_cfg() and scheduler.step().

Alternatives

• Single-request execution (Current): Requests are processed sequentially (execute_request), leading to GPU under-utilization during offline inference benchmarks or multi-prompt batch jobs.
• Step-wise execution: Requires implementing the full SupportsStepExecution state protocol (prepare_encode, denoise_step, step_scheduler, post_decode), which involves a larger architectural refactor. Request-level batching provides a simpler and faster path to multi-request inference.

Additional context

• Reference implementations in vLLM-Omni: LTX23Pipeline (vllm_omni/diffusion/models/ltx2/pipeline_ltx2_3.py) and QwenImagePipeline (vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image.py).
• Pipelines affected: Wan22Pipeline (vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py), Wan22I2VPipeline (vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_i2v.py), Wan22VACEPipeline (vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_vace.py), and Wan22S2VPipeline (vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_s2v.py).