`PipeOpTorchModel` should expose fields `callbacks`, `optimizer` and `loss` so they can be changed afterwards

Also requires: https://github.com/mlr-org/mlr3pipelines/issues/888


This would enable:


```{r}
graph = po("torch_ingress_num") %>>%
  po("nn_head") %>>%
  ...

graph$configure(
  torch_model_classif.callbacks = t_clbk("history")
)
```