Adjusted ActNorm to work as described in the paper

[LarsKue]

This is a follow-up PR from the discussion sparked in #165.

This PR involves:

• Initializing the ActNorm Parameters to unit size in all dimensions except the channel dimension
• Drawing the mean and std over the batch dimension, as well as any dimensions beyond the channel dimension
• Checking the batch for correct shape before overwriting the parameters
• Adjusting the tests to reflect these changes

For free with these changes comes:

• A fix for silent device moves in the initialization method
• An additional check for an initialization batch that has zero standard deviation in any channel

[LarsKue]

Please do not yet merge, there appears to be an error with the log jacobian in another test. EDIT: Resolved.

[fdraxler]

Looks good

[RussellALA]

Looks good to me, I think you can merge