Refactor Configs

[mrwyattii]

In order to better support multi-model deployments and associated features from #208 we have identified some key ways in which the configs can be refactored. This will be the first of 2-3 PRs that does some refactor work and implements smaller segments of the work done in #208 to make review and merging easier.

@TosinSeg

Work done in this PR:

• Refactor our configs to have MIIConfig and DeploymentConfig. This is in anticipation of having multi-model deployments where we will be passing in multiple DeploymentConfig objects
  • This could probably be extended in this PR to also include the DeepSpeed InferenceConfig rather than recreating that config in load_models.py (see comments I left on the code changes)
• With the refactoring of our configs, we can simplify a lot of our code:
  • Avoid packing/unpacking config data at various stages of the deployment. Instead, just pass around the mii_config object as it contains everything we want. This massively simplifies the score template and server class
  • Move lots of parameter and config verification that was scattered around the codebase into the config classes themselves. This consolidates code to a common location and simplifies the process of debugging config settings
• Light formatting changes. We use yapf for formatting, but this allows mixed ' and " - I don't think we need to use one or the other, but being consistent is good. Applied black and then yapf to get some nicer formatting (especially on the long assert lines)
  • Is it too late to switch the project to use black? @jeffra

I left lots of comments about changes I felt unsure about or where I could use some feedback, however the PR is mostly ready to merge. We could always handle minor cleanup and changes in future PRs that implement the rest of #208

[tohtana]

Thank you for this great effort! I've noticed that the areas I had concerns about have been improved. This will make future development much more pleasant.