More Multimodal RL support

Thanks for your wonderful job of the RL framework ROLL.

Do you have plans to support more multimodal RL pipelines, including audio generation for some omni models and even some image/video generation models? And for inference, vllm-omni and sglang diffusion are also pushing to serve these multimodal models.

The issue names no files, tests, or entry points. Begin by reviewing ROLL's current multimodal RL and inference support, then assess the requested audio, image, and video pipelines alongside vllm-omni and sglang diffusion; done would require a defined implementation scope rather than a general roadmap request.

Written by the indexing model from the issue text.