Revise capability hierarchy and fix classifiers

[odersky]

We now make Capability a sealed trait with two subtraits: SharedCapability and ExclusiveCapability. This forces one to pick one or the other when defining a new Capability type. Most tests were changed to use SharedCapability instead of Capability. Since classifiers are now much more common than before, this showed up some bugs and anomalies for classifyer handling, which are also addressed in this PR.

Details:

1. Further improvement to error messages. We now explain why FreshCaps cannot subsume a capability.
2. We turn caching off by default for captureSetOfInfo. During experimentation I found that sometimes captureSetOfInfo could be cached too early, leading to a locked in universal cap in the info where we would expect to see a fresh cap. On the standard library it did not look like this had a noticeable performance impact.
3. Fixes to tryClassifyAs and transClassifiers. In particular, we need to distinguish FreshCaps that can still be classified from ones that cannot. Once a FreshCap is part of a constant capture set, it gets classified by the type that prefixes the set and that classification cannot be changed anymore. But other FreshCaps are created as members of variable sets and then their classification status is open and can be constrained further.
4. Convert a fresh cap to a reach capability only if the reach capability might subcapture the fresh cap. Since fresh caps are now classified that is not always the case.
5. Turn pre-type closure results on for parametric expected function types. This gives sometimes better errors and avoids premature mapings to scoped result types in closure bodies.
6. Use cap.rd as implicitly added capability only for references extending ExclusiveCapability. Use cap for the others.

We also include now captures in the result type of a function in the captureSetofInfo of that function.

[odersky]

I think this is ready for review now.

[Linyxus]

LGTM!