add decomposition for nll_loss2d_backward (#77198)
Adds a decomposition for `nll_loss2d_backward`
This will let us actually run all the tests for jvpvjp ([see this functorch PR](https://github.com/pytorch/functorch/pull/792)). I confirmed locally that this made those tests pass too
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77198
Approved by: https://github.com/Chillee