[Cringe] edits to metric logging

[jxmsML]

Patch description
Main change:

1. target_token
   before: target_tokens = notnull.long().sum()
   after:

        notnull = batch.label_vec.ne(self.NULL_IDX)
        ce_mask = torch.logical_and(notnull, ce_mask.view(-1, batch.label_vec.size(-1)))
        ct_mask = torch.logical_and(notnull, ct_mask.view(-1, batch.label_vec.size(-1)))
        # number of tokens in each examples for cross entropy or cringe loss.
        metric_notnull = torch.logical_or(ce_mask, ct_mask)
        target_tokens = metric_notnull.long().sum(dim=-1)

2. logging other
   metrics:

1. ppl_ce: ppl on tokens computed for lm nll loss
2. ce_target_tokens: count of tokens computed for lm nll loss
3. ct_target_tokens: count of tokens computed for cringe loss

Testing steps

Other information