Do two passes through delayed elaborators (it's impossible to predict
dependency order). Possibly this should keep going as long as it's
making progress? I'll consider that later.
We can work out what it should be since each case is supposed to have
the same type, so we can look ahead to the alternatives. This way, we
don't have to rely on the scrutinee alone to calculate the type of the
case block. (Idris 1 does this too but I've only just encountered the
need for it in Idris 2 now!)
