introduce transforms parameter that wraps data and targets

[biphasic]

Sometimes you might want to apply a stochastic transform to data and targets together, i.e. flip the events horizontally with a given probability and if it is flipped, then also modify the label accordingly. This is currently not possible using the separated calls of transform and target_transform. PyTorch Vision adds (for some datasets) a third parameter transforms which wraps both data and targets.

[biphasic]

done in 55709c6