TRL Reduces Agentic RL Trainer Compute 26x via Tree Packing

TRL's AsyncGRPO implementation now uses a tree packing mechanism that forwards each unique token once per batch instead of repeating shared rollout history, allowing trainers to reach the same reward on agentic reinforcement learning with 26x less compute on MiMo Envs. The update replaces an earlier FlexAttention-based iteration with a single FA3 varlen call and addresses bottlenecks in forward and backward passes when agentic training rollouts exceed 100k tokens.