We currently only show the first error seen. Using a different error applicative functor, we can accumulate a set of errors. We need to be careful about this because it is not appropriate everywhere, but we can, for example:

1. collect all name resolution errors
2. collect at most one error from each checked type