Add support for context-aware revisions

[miltondp]

The current strategy revises one paragraph at a time independently. That is one independent query per paragraph. This loses context. A way to add support for context-aware revision is to use the ChatCompletions endpoint by including a "system" role and the history of revisions.

[miltondp]

An alternative or complement to this is to use models with large context, such as the new gpt-3.5-turbo-16k