Learning Adaptive Parallel Reasoning with Language Models