Towards Robust and Fair Next Visit Diagnosis Prediction under Noisy Clinical Notes with Large Language Models