Adding support for patching

[aspctu]

Overview

This PR adds support to Truss to apply monkeypatches to various libs. It also adds a new field in the config enable_patches (defaults to True) which enables/disables this behavior.

This PR also includes 2 patches:

1. Huggingface patch to parallelize model downloads
2. Whisper patch to make Whisper downloads b10cp compatible

Testing

I tested the following flows

1. Local server flow with the flag enabled / disabled
2. Docker image flow with the flag enabled (this requires building a new base image, instructions below).

To test the docker flow, following the steps below:

1. You'll need to create a new base image using the helper script in bin/generate_base_images.py. To conserve space on your device (this script can create many images at once), you can run the following command to produce a non-GPU, non-live reload, serving image:

python bin/generate_base_images.py --version-tag 5 --job-type server --use-gpu n --live-reload n --python-version "3.9" 

2. Once you have this image, you'll create an iPython shell and run the following commands

import truss
tr = truss.load("path to truss with huggingface / whisper model to test patches")
tr.set_base_image("baseten/truss-server-base:3.9-5")

tr.docker_predict({"prompt" : "man on moon"})

[pankajroark]

Would be great to have at least some unit tests for the patches.

[bolasim]

@ujjwal @pankajroark I addressed all comments with the exception of the unit tests for the specific patches. Don't want to get blocked on that for rolling this out. Let's do that in a followup.

Let me know if there is anything else that needs to get addressed here before rollout

[pankajroark]

We should double check that these logs don't show up. I think it will be common that not all patches would apply for a model.