DataTransformer init parameters

[FlorentRamb]

This PR solve issue #7, it allows two things:

1. ability to fit gaussian mixtures on a subsample (help to scale with big datasets while losing only little accuracy)
2. ability to pass init arguments to the DataTransformer through CTGANSynthesizer.fit (and so to change other parameters as max_clusters).

[CLAassistant]

All committers have signed the CLA.

[candalfigomoro]

@fealho @pvk-developer Can we merge this? It's basically impossible to fit the CTGAN on a large dataset because the gaussian mixture is a huge bottleneck (even using dozens of CPUs). This PR would allow to speedup the gaussian mixture step. Thanks

[fealho]

@npatki not sure what you want to do with this?

[candalfigomoro]

Meanwhile the library code has changed so the PR should be updated.

For example, the _fit_continuous method now receives a pandas DataFrame, so np.random.choice() can be replaced by something like data = data.sample(self._max_gm_samples, replace=False, random_state=SEED).

Also, I wonder if ClusterBasedNormalizer could not be optionally replaced by a power transform, which might be faster (although it might impact the quality of the generated data), see https://github.com/sdv-dev/RDT/issues/613