Do you really need to pretrain Q-functions for online RL fine-tuning?