Avoid setting transforms in the buffer

[lballes]

Currently, the DataBuffer receives transform and target_transform and applies them when accessing an item. When deserializing the Learner state, this requires resetting the transforms (see DataBuffer.set_transforms) and generally bloats the code of the buffer.

I would propose to remove the transforms form the buffer and instead generalize the _TransformedDataset such that it can be applied to DataBuffers as well. This way, transforms are handled in a unified way by wrapping all relevant datasets in _TransformedDataset in the Learner.

To make _TransformedDataset more general, instead of passing transform and target_transform, we could pass a single transform that operates on the entire data point. When wrapping a buffer in _TransformedDataset, we just need to make sure that the metadata are "ignored" (e.g., transform=lambda (data_point, metadata): (my_transform(data_point), metadata)).

[lballes]

Any concerns on this? Otherwise I would add this to the next milestone and take care of it.