Refactor logic such that all AI message handling is provider-dependent

Due to subtleties in each provider, this leads to fragile, difficult to understand code that has led to strange edge case behaviour for our users.

We need to move towards having all message handling be provider-dependent in order to make it easier for us to maintain and extend these integrations going forward.

Example:

* `truncate_and_annotate_messages`
* Message truncation will eventually be going away so when that happens, it'll be a good opportunity to deal with this function