Language-Conditioned Representations and Mixture-of-Experts Policy for Robust Multi-Task Robotic Manipulation