Fixing create_tfrecords.py when using custom tokenizer

[galloj]

Fixing error when loading tokenizer using encoder_path. Also I added some error messages, because all errors which happened because of bad input were mostly unreadable.

I think it would be nice to do some more refactoring on this file too, because I think there will be bad handling of errors in other cases too. Also I had issue when multiple threads were used, then multiple tensorflow instances were initialized at once, which caused that both of them printed at same time, which resulted in letters being printed in random order.

[StellaAthena]

This looks reasonable, but I haven't been able to test it yet. Just wanted to let you know that I am aware of this PR and it's on the to-do list.