floq: Training Critics via Flow-Matching for Scaling Compute in Value-Based RL