[Feature]: enhance diffusion dynamic batching scope

[asukaqaq-s]

🚀 The feature, motivation and pitch

related to #874

Current dynamic batching for diffusion only supports requests with the same resolution. As a result, when benchmarking on diffusion dataset C, many requests with different resolutions still cannot be batched together, causing low MFU utilization on a single instance.

The goal is to relax the constraints inside SamplingParamsKey and allow more heterogeneous requests to share one forward pass.

1. Remove CFG-related constraints:

   • When all CFG scales are != 1, different requests can still be batched into the same forward pass.

   • Additional considerations:

     • whether positive/negative prompts should be batched together
     • how to support CFG parallelism efficiently

2. Remove height/width/frame constraints:

   • Support mixed-resolution batching with FlashAttention varlen functions.
   • Pass attention metadata through the existing attn_metadata interface.

Alternatives

#3638
MixFusion is relatively complex, so we should also compare it with the simpler varlen batching approach and evaluate which one provides better performance gains.

Additional context

Test plan:

• Single-GPU correctness and performance validation
• Support CFG parallelism
• Support Ring Attention / Ulysses Attention with padding-free execution
• Optimize scheduling between StagePool and the diffusion scheduler

@hsliuustc0106 @Semmer2 @wtomin @wuhang2014 @Gaohan123 @cherhh

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.