Refactor models.py

[LuxF3rre]

• Removed unused imports
• Formatted with black
• Extracted tokenizers and model loaders into their own functions
• Changed model parameter calling to use kwargs, extracted CPU failover and suggested memory allocation into helper functions, minor readability improvements for memory_map allocation.
• The general logic untouched

[oobabooga]

I find that this makes the code significantly more verbose and harder to read.

Model loading with *kwargs has been also handled in #358