The original variance work only used variances in subtyping. However,
there are a couple other places that variance is relevant that this
adds support for.

- When solving a variable to a type involving abilities, we relax
  ability lists in strictly positive position with an extra ability
  variable if they don't have one already. This allows for e.g.
  types with different concrete ability lists in separate branches
  of an if/case to resolve to the combined list.
- When generalizing, we discard variables that occur only in covariant
  position. This avoids some cases that build up multiple ability
  variables in a single list, which confuses the type checker. It
  also generally produces cleaner looking types.

Previously these only applied to arrow types, but now it also works
based on inferred variances.

…vements