Since the NF might refer to hole names, and those hole names might be
possible to evaluate now, we'll need to recalculate the expected type's
normal form before rerunning the delayed elaborator
Works by running all possible elaborators and checking that exactly one
succeeds. Still to do: pruning the list of elaborators by target type,
dealing with 'UniqueDefault', checking that delaying on failure works as
it should.
When we encounter them, not that they're a binding as normal, but also
record the thing they expand to. Then bind as a PLet, and convert that
to a Let on the RHS so it has computational force. The case tree
compiler knows about as patterns, so they get compiled to use the
appropriate name on the rhs (rather than a let).
