Scalable Training of Mixture-of-Experts Models with Megatron Core