Transformers library refactors MoE weight loading for improved performance

AI Impact Summary

Mixture of Experts (MoEs) are gaining traction in large language models due to their ability to achieve better compute efficiency and scaling compared to traditional dense models. This change introduces a significant refactor to the `transformers` library, specifically around weight loading, to handle the serialized nature of MoE checkpoints. The new WeightConverter and lazy materialization techniques dramatically improve loading speeds, addressing a key bottleneck in training and inference with these models.