CTM Memory and Speed Improvements

[Supermaxman]

I regularly ran out of memory on large datasets during the CTM fit call. Upon further inspection, I found the automatic generation of training_doc_topic_distributions with get_doc_topic_distribution after the fit call to have many opportunities for improved memory usage. I made the following changes:

1. Made generating the training_doc_topic_distributions optional at the end of the fit call, defaulting to True
2. Removed the n_samples loop from the get_doc_topic_distribution operation
3. Reused the posterior_mu, posterior_log_sigma across samples, as they remain the same
4. Vectorized the sampling process on GPU
5. Added a get_doc_topic_distribution_iterator operation, which significantly saves memory if the dataset is large by computing topics one batch at a time and yielding them as an iterator
6. Formatted the ctm.py and decoding_network.py with black, for easier visibility

I do not believe any of the changes I made are breaking changes, these should simply make topic discovery run faster and with less memory.

These efficiency gains enabled me to run the CTM on a massive collection of tweets, so I thought I would create a pull request and offer these improvements back to the original repo, as it still seems active.

[vinid]

thank you so much for this! I'll go over the changes hopefully tomorrow and merge them!

[vinid]

renamed log sigma to log var. I know log sigma is more coherent than var but the variable is referred as log_var in the rest of the code; i'd keep it like this for consistency.

[vinid]

merged!