Less Context, Same Performance: A RAG Framework for Resource-Efficient LLM-Based Clinical NLP