Towards Adaptive Humanoid Control via Multi-Behavior Distillation and Reinforced Fine-Tuning