Investigate whether we can support simplified and deep subsumption

In the call today we discussed the reddit thread which discussed the simplified subsumption proposal. And this Haskell Discourse thread

@simonpj indicated that it might be possible to proceed in a manner similar to let generalisation where there is a predictable/simple subsumption mode (ie simplified subsumption) but also unpredictable/deep subsumption mode (ie, the old behaviour).

This ticket is created to track that idea, which Simon wanted to discuss with @rae.

This is perhaps important for tickets like #20818