Rethinking Test-Time Scaling for Medical AI: Model and Task-Aware Strategies for LLMs and VLMs