Reducto trained its R1 model using reinforcement learning loops around mid-training and post-training on edge cases