Bugfix/rand split torch data

[zeyus]

I'm using the OpenBCI cyton/daisy setup, and it doesn't have EOG channels setup by default, also we're not working with huge data sets so we want to concatenate the data from one participant but split it into sets. As such this PR is to:

• Added MNE related options to conf
  • notch filtering
  • adding an average ref channel
• merge PR Should fix #73. #75 and bugfixes #76
• Fix Thinker.split() returning unwrapped torch.utils.data.dataset.Subset instead returning a new DN3ataSubSet object which wrapps Subset and includes a get_targets method to allow using the returned training, validating and testing variables inside a BaseProcess.fit method with a balance_method argument