Reset loss to zero on logging in Trainer to avoid bfloat16 issues (#8561)
* make tr_loss regular float
* Revert "make tr_loss regular float"
This reverts commit c9d7ccfaf0c4387187b0841694f01ec0ffd5f4ba.
* reset loss at each logging step
* keep track of total loss with _total_loss_scalar
* add remaining tr_loss at the end