Evaluating Retrieval-Augmented Generation vs. Long-Context Input for Clinical Reasoning over EHRs