Event weighting for imbalanced datasets

Thanks for the great library! I'm training with an imbalanced dataset. In time-to-event prediction, is it reasonable to weigh positive examples, as can be done in the classification analog? If so, could a convenience keyword arg to the loss functions be added (see `pos_weight` [here](https://pytorch.org/docs/stable/generated/torch.nn.BCEWithLogitsLoss.html))? I think this would be similar to, but more efficient than, oversampling positive examples.

Start by reviewing the repository's existing loss functions and the PyTorch BCEWithLogitsLoss documentation for `pos_weight`. Determine how event weighting should apply across the survival losses and how it should compare with positive-example oversampling; the work is done when the supported keyword behavior is defined and consistently implemented.

Written by the indexing model from the issue text.