I tried to extend specialization to all context functions, not just ones of 0 arity.
But that runs into problems for dependent context functions, since the necessary casts
get complicated. Since context functions over primitive types are an anti-pattern anyway
I don't think we need to optimize this case, after all.

The regular changeOwner can cause cycles in transformations. Transformations like ByNameLambda should use
the less demanding changeOwnerAfter.

There was an omission before, but as long as  byname parameters were ExprTypes it did
not lead to problems.

Instead, drop <by-name> applications when pickling and reconstitute them
based on formal parameter types when unpickling.

ByNameLambda needs to run in a group after ElimRepeated since ElimRepeated
works on ByName arguments but not converted closures, and it sees the arguments
after transformations by subsequent miniphases in the same group.

ByName nodes in arguments are not pickled, which means that we
can use the same Tasty version as before.

 - Handle combinations of NamedArg and ByName
 - Accept and evaluate any remainign ByName applications