Add gpu support

[Kostis-S-Z]

What's changing

Add support for both the text-to-text and the text-to-speech model to be loaded on the GPU.

How to test it

Steps to test the changes:

1. Run the demo
2. Check with nvidia-smi that your GPU has loaded the models

Additional notes for reviewers

This became more difficult than I expected because the two different frameworks of the models need to have support for the same cuda toolkit. Support for the text-to-speech is complete and was easy. Support for the text-to-text has proven quite difficult.

Some rough benchmarks:
Setup: Count time from the moment you upload the document until the first audio sample of speaker 1 is generated (includes loading both models and running inference once with both models). The .html file in example_data was used. GPU: RTX 2060

CPU: text-to-text & text-to-speech -> 2min, 41sec
CPU: text-to-text GPU: text-to-speech -> 1min, 27 sec (!)

I already...

• Tested the changes in a working environment to ensure they work as expected
• Added some tests for any new functionality
• Updated the documentation (both comments in code and under /docs)