[train] try out vl norm loss aggregation method

[justinvyu]

We currently support a few loss reduction methods:

1. sequence_mean: GRPO
2. token_mean: DAPO
3. seq_mean_token_sum_norm: Dr. GRPO

This paper proposes a different way ("VL Norm") to reduce the loss that "provides an unbiased estimate of the true policy loss but also minimizes gradient variance"

cc @erictang000