Remove by-mut-ref mode altogether

[nikomatsakis]

The handling of "by mutable reference" parameters is unsound because it doesn't affect variance. This is not news. One place we take advantage of this is vec::push(), which has the type

fn push<T>(&vec: ~[const T], +val: T) { ... }

This is unsound because we accept any sort of vector for the argument vec, but the type system only guarantees that we'll write back a const vector. I thought this was basically harmless, since vectors are uniques, but I forgot that it also implies that we will be covariant with respect to T. When combined with regions, this leads to bugs like #3501. However you can create problems without regions too. Any place where we have subtyping.

Anyway, we had always planned to remove by-mut-ref mode as part of the general "de-moding", but I think we should up the priority for this particular mode. It is not that widely used in any case, though vec::push() is certainly frequent. I am checking out how hard it will be to just purge it altogether.

[graydon]

Yeah, this is the really major one.

[catamorphism]

I'm not sure if we decided this was for 0.4 or not, but I'll try to do it.

[catamorphism]

I am working on this, but ran into a strange bug where apparently, the code after a nested pure fn gets treated as if it should be pure. I'll try to isolate it tomorrow.