large-rollout

Introduces DualKV, a FlashAttention kernel variant that eliminates redundant prompt token computation in RL post-training (GRPO/DAPO), achieving up to 3.82x speedup on 30B MoE models.