Support LLM-based Merging

[drewmccormack]

With a 3-way merge, there is presumably enough information for a LLM, even a small one, to come up with a good merge of divergent text. This seems like a natural evolution of data models: rather than using traditional algorithms, when a conflict arises on complex data like human-readable text, an LLM would presumably merge it in a more natural way.

CRDTs are partially fulfilling this role now in the framework, ie, making more "human" choices when merging text. But even those are just an algorithm attempting to do what a human might. An LLM is a much more advanced algorithm which is trained at exactly this type of problem, and an embedded LLM on a device should be able to do a great job of merging conflicts in text.

This would be a good issue for someone who has some experience with embedded LLMs. The idea would be to add a new type of merger called a LLMMerger. For inspiration, see TextMerger. The macros would then be updated to support a new merge algorithm, eg, @Merged(using: .smartTextMerge).