MIR: Involve Micheline in the core data model

This is the follow up issue for this discussion.

Specifically we should involve Micheline in places where it makes sense to do so. At this point, the most compelling reason appear to be the implementation of PACK/UNPACK operations who's semantics seemed to be closely associated to Micheline, so it seems that we will have a more uniform way of implementing it, if we go via Micheline.