* Normalise types fully at the REPL
It was a bit odd that we only normalised the scope of function types and
not the arguments, and I can't remember the reason for that if there
even was one.
* Better way of using nf_metavars_threshold
If a term is getting big and probably needs normalising, we now have a
sizeLimit flag in quote, so we can use that instead of checking the size
afterwards. This is a handy heuristic for speeding up unification when
there's a term with lots of suspended computation. Fixes#1991
