Avoid necessaryEither because it could constrain too much

[odersky]

In a comparison A & B <: C { x: D, y: E }

we cannot reduce the problem in a sound way to necessaryEither. Example:

A = C { x: D }
B is a constrainable type variable

Here, the strongest constraint that follows from A & B <: C { x: D, y: E } is

B <: C { y: E }

But necessaryEither would give in this case

B <: C { x: D, y: E }

Fixes #12306

But a better fix could be to decompose the refinement type on the right into its constituents and
run a necessaryEither for both A & B <: C { x: D } and A & B <: C { y: E }

[odersky]

It turns out that the problem is specific to refinements on the right. It does not come up with AndType/AndType comparisons. So
#12317 is a superior fix.