Add support for half precision in face renderer

[vardanagarwal]

By adding a support for half precision, I was able to speed up the operations by more than 50%.

To use in the inference.py add the --half flag. In the gradio demo, created a checkbox for enabling half precision.

Initially, I was applying it on all the models in src/modules/make_animation, but then the output head movement was almost non-existant so sticking to applying it only on the main generator. The output is also very similar to the one without half precision but not entirely same.

I have tested it with different combinations of preprocessors and reference videos and did not face an issue, but if there is some test suite to test against please do that or let me know.

[vardanagarwal]

#625 #309

[2catycm]

Awesome! I am trying to make sadtalker faster too! This PR helps me a lot