Fix inverse token mask concatenation error and add support for mps devices

[EigenTom]

Thank you for your great work and generous contribution. During my reimplementation of your work I have noticed the following error:

in /bert/dataset.py, at line 197 in method _preprocess_sentence(), the default value of inverse_token_mask has been set to None, which cannot be directly appended to a List object. With Python 3.11, running the code will result in the following error:

which can be resolved by replacing the variable's default value as an empty list [].

Having noticed you haven't enabled mps device acceleration support for Apple Silicon devices, I have added support for it as well. It will make full use of Apple Silicon SoCs' GPU for model training and should yields much faster model training speed.