[dtensor] add op support for nll_loss_backward (#119256)
As titled. This is a followup to PR #118917 on nll_loss_forward. It also fixes an issue in it: the forward function produces two return values, the loss `result` and the `total_weight`. The previous PR didn't explicitly deal with the `total_weight` part.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119256
Approved by: https://github.com/wanchaol
