Fix #18246: correctly compute capture sets in TypeComparer.glb

[Linyxus]

Intuitively, glb(S1^C1, S2^C2) should capture C1 ⊓ C2, where ⊓ denotes the glb of capture sets. When computing glb(tp1, tp2), if tp1 matches CapturingType(parent1, refs1), it captures C0 ++ refs1 where C0 is the nested capture sets inside parent1. Denote the capture set of tp2 with C2, the result should capture (C0 ++ refs1) ⊓ C2. Presumably, ⊓ distributes over the union (++); therefore (C0 ++ refs1) ⊓ C2 equals C0 ⊓ C2 ++ refs1 ⊓ C2.

Previously, if C2 is a subcapture of refs1, it is simply dropped. However, based on the above reasoning, we have refs1 ⊓ C2 equals C2; therefore C0 ⊓ C2 ++ refs1 ⊓ C2 equals C0 ⊓ C2 ++ C2. C2 should not be dropped, but rather kept. This PR fixes the logic to behave correctly.

Fixes #18246.

[odersky]

Denote the capture set of tp2 with C2, the result should capture (C0 ++ refs1) ⊔ C2.

I guess that should be \squarecap not \squarecup.

[odersky]

Nice catch!

[odersky]

Maybe we should roll the isAlwaysEmpty optimization into capturing?