Consider moving training related parameters to `fit()` about ctgan

[csala]

Thanks for the suggestion, @kevinykuo , but I'm afraid I disagree with this.

The rationale behind the parameter organization is:

• __init__ arguments specify "how" this instance is going to behave, i.e. "how" this model will learn.
• fit arguments specify "on which data".

The only argument that could be moved to the fit method is epochs, and only the code is changed so the creation of the internal instances is not done every time the fit method is executed, which would then allow resuming a previous fitting process (see #4 )

All the other arguments are fine where they are and should be kept there.

[kevinykuo]

Thanks @csala! I think your comments make sense. A couple of use cases (neither of which hits me currently) that may benefit from moving batch_size are 1) one may want to continue training a model on a different machine that doesn't enough GPU memory to support the batch size specified and 2) adaptive batch sizes. This is pretty minor though, so no hearts be broken by a wontfix :)