Refactor Input Transforms in RL

[fabioseel]

Currently, input transforms are implemented via hijacking the "input normalization" of samplefactory.
However, this is a weird hack and confusing.
There's two better ways of doing it:

Make the transforms part of the model

The transforms are torch modules, so they can easily be added to be part of the model.
Advantages

• simple solution (actually similar to the current one, the input transform is part of the model, but a bit cleaner)
• easy access to 'untransformed' input eg for denoising

Problems

• logically, they are not part of the model
• reusing a model seems to be slightly more difficult (either the model config but especially loading / storing the weights needs some extra handling in that case)

Use environment wrappers

One could also use an environment wrapper to implement the transforms, pretty much as in #84
Advantages:

• logic-wise / conceptually the cleanest solution

Problems:

• moves the computation to the 'environment' processing, hence cpu, which already seems to be the bottleneck
• need to explore how easy it is to access the 'untransformed' input in that case

We should decide and implement this before starting to work on #87 .