dataset transform and target_transform

[vincentqb]

This is a spin-off discussion from comment about the addition of transform and target_transform parameter to dataset (cc @fmassa for torchvision mentioned in comment). They were removed in #303 to simplify the interface.

What are options?

• transform and target_transform for every supervised dataset, and transform for unsupervised?
• aProcessed class as introduced in Example pipeline with wav2letter #632
• pre-processing in the collate function