Add load balance loss

[RissyRan]

Description

• add load balance loss to MoE dropping strategy.
• set default load_balance_loss_weight as 0.01 (from paper that this value loads balance quickly without interfering the training loss)

Test

local test on a small model size:

• training 100 steps without lb_loss (capacity_factor=-1): baseline - lb_loss=0
• training 100 steps with lb_loss (capacity_factor=1): test