Make Word2vec multiple passes more explicit. Fix #1052

[robotcator]

I modify the method 'train()' by adding some parameters --iteration=None, epochs=1, start_alpha=None, end_alpha=None, cur_epoch=None. And the default epochs is one, others are None. If the user trains the model with a loop, the epochs and cur_epoch(the current epoch) must be specified. So if the user don't specify the start_alpha and end_alpha(the usage is the same as alpha and min_alpha), we can recaculate the start_alpha and end_alpha according to the current epoch to support linear learning-rate decay from (initial) alpha to min_alpha.

[robotcator]

I also have another idea that we can force the user to pass those parameters(iter, alpha, min_alpha) explicitly. If the user train the model using a loop, they can carefully pass the start_alpha and end_alpha to support one long smooth descent from alpha to min_alpha.

[gojomo]

Another pending PR with same general idea is #1139 - forcing users to be aware of train()'s parameters by making them required.

[tmylk]

@robotcator Thanks for this contribution but it is best to close this PR and finish #1139 in particular add tests for assertRaises if the new epochs param is ommiteed; update the ipynbs that use train; implement epoch_callback,

[robotcator]

@tmylk Thank you for your suggestion. But can you give more details about implementing epoch_callback ?
@gojomo Would you mind let me to finish the #1139 to do list.

[gojomo]

@robotcator - Please feel free to continue/finish #1139. In my opinion the most important thing is to ensure notebooks/docs reflect the change & current best-practices, and resolve any conflicts with the main develop branch.

Implementing epoch_callback would require some careful refactoring of the existing loops and multi-threading; I'm not sure of the best way to proceed yet, which is one of the reasons I put #1139 aside.