Add support for keras models

[juliasilge]

Addresses #144

[juliasilge]

Thank you again for this review @t-kalinowski! I outlined some of where we are and where we might head from here in #55. To sum up:

• This now errors on creating the vetiver_model() object for multi-input keras models, because the current strategy for the POST endpoint will not handle that well. Let's hear from users about how important this is.
• The default behavior is to require the user to pass in some prototype_data (this is similar to ranger and all of the Python implementation) or to explicitly turn off checking of new data with save_prototype = FALSE. If someone wants to build a model with non-rectangular data like images, this is currently what they will have to do. Again, let's hear from users about how important this is and what they want to do.

[juliasilge]

I have made some good headway on how to deploy this to Connect today. Some notes:

• The automatic dependency discovery for Python didn't work for me at all, I think because I am on an M1 and all my TensorFlow stuff is specific for ARM architecture. I could solve this by deploying the plumber file together with a minimal requirements.txt:

keras
tensorflow

I did successfully deploy a model! 🎉

• Sadly, the model can't predict. 😞 From the logs, I suspect that the problem is that I trained and serialized the model on my M1. For the next step, I am going to try to train the model on Colorado Workbench and then deploy from there. I'm currently having trouble installing TensorFlow on Workbench, however.

[juliasilge]

Still TODO: create a minimal requirements.txt to deploy with the plumber.R for Connect.

[juliasilge]

This is working now on Connect, I believe!
https://colorado.posit.co/rsc/content/9c37e89e-c094-41b8-813c-2f636dd98e51

I will work on a Docker container in a different PR.