Add feature that allows to train a `LearnerTorch` on task where the target is also a lazy tensor

[sebffischer]

The general approach would be to:

• Define a TaskClassifTorch where the target feature type is a lazy_tensor (a long)
• Define a TaskRegrTorch where the target feature type is a lazy_tensor (float)

Because we don't want to reimplement everything (measures, learners), we need a way to convert this to a TaskClassif and TaskRegr respectively.

For this we need a custom DataBackend that returns the target as a numeric (regr) or factor (classif).
However, this DataBackendTorch should also have a method to directly retrieve the underlying tensor data which will be used by the LearnerTorch when iterating the batches.

We then need converters

as_task_classif.TaskClassifTorch
as_task_classif.TaskRegrTorch

The whole code for e.g. using torchvision::dataset_mnist() would then be:

ds = dataset_mnist()
task = as_task_classif(ds, target = "y", input_shapes = list(...))
learner = lrn("classif.alexnet")
learner$train(task)