Support for -instruct models

[machinewrapped]

I added support for the GPT-instruct models that OpenAI have exposed to the API, which use a different API method. This required/encouraged refactoring the system so that the API-specific calls are encapsulated, and the translation process is more or less agnostic about the particular model to be used for translation.

It's quite a big refactor, which will likely be a pain to merge into any forks, so I open it up to discussion prior to integrating the branch back to master:

#86