Date, IDate, and POSIXct should be distinct classes for joins

[MichaelChirico]

I think we should implement mergeType returning I?Date, POSIXct as separate cases. That would for example solve #3506 and could make the solution here clearer.

Originally posted by @MichaelChirico in #6603 (comment)

#6602 highlights difficulty with our current approach of ignoring column class attributes for well-known classes. We already cater for two very common classes (factor and integer64); I see no reason to treat Date, POSIXct any differently.

Solving this would also solve #3506 with more generality. We should also keep #6627 (in particular the linked "join compatibility table") in mind.