Xiangyu Li

[megatron] fix: expose moe_aux_loss_coeff and moe_z_loss_coeff to improve MoE load balancing

2

### What does this PR do? When training MoE-family models with Megatron as the backend, enabling Expert Parallelism (EP) may cause load imbalance across experts, which makes the update_actor step...