Roadmap for adding vision language model support

[Jackmin801]

Add comprehensive analysis document outlining the changes required to support vision language models like Qwen3-VL-8B-Instruct in Prime-RL's SFT and RL training pipelines.

Key areas covered:

• Current architecture overview and component interactions
• Qwen3-VL model architecture requirements
• Core changes needed for model loading, forward pass, and data structures
• SFT and RL pipeline modifications
• Inference server integration with vLLM
• FSDP sharding for vision encoders
• Implementation roadmap with phased approach
• Technical considerations for memory, batching, and serialization