plotting and exhaustive saturation computation

Reworked the interface of saturation:
** it is now possible to pass kwargs over to the writer object
** it is now possible to pass writer objects directly instead of a string key
Added another writer which saves saturation plots additionally to the csv-file.

multi-gpu support

Minor changes that enabled saturation to be computed on multiple GPUs.

Maximum Sample Feature

enabled logging with limited number of samples. This new feature is off by default and can be enabled using "max_samples"-parameter in the CheckLayerSat constructor.
This is usefull to gain performance if batch-processing is fast on it's own.

fixed sample logging

enabled logging with limited number of samples. This new feature is off by default and can be enabled using "max_samples"-parameter in the CheckLayerSat constructor

Covariance Reset Computation

Added a switch in the constructor for resetting the covariance matrix after each '''add_saturation()''' call.

hotfix

minor hotfix

minor bugfix

fixed a bug where add_saturations() would sometimes couse a crash.

Merge pull request #17 from delve-team/feature/pytorch_refactoring

added version.py

minor bugfix

fixed a bug where sometimes add_saturations would crash

refactored torchcallback

removed most of the unused functionality
added different types of logging (csv, console-out are now alternatives to tensorboard)
fixed a bug where the system would not search properly through arbitrary nested modules.
changed the constructor interface of CheckLayerSat() in order to be more ergonomic in use
changed the general interface of the functions to be more closely related to Sacred and MLFlow-Type interfaces.