Closes #7331.

Motivating example:

trait FromVec<'self> {
    fn fromVec(&'self [u8]) -> Self;
}

struct Reader<'self> {
    v : &'self [u8]
}

impl <'self> FromVec<'self> for Reader<'self> {
    fn fromVec(v : &'self [u8]) -> Reader<'self> {
        Reader::<'self> { v : v }
    }
}

impl <'self, T : FromVec<'self>> Reader<'self> {
    fn get(self) -> T {
        FromVec::fromVec(self.v);
    }
}

When we typecheck FromVec::fromVec, we get fn(&'self [u8]) -> T where T is bounded by FromVec<'self>. Usually we would replace the argument's 'self with a new region variable which could match anything. However, if we allow that here then we lose an important constraint between the argument and the return type T, namely that the 'self in FromVec<'self> needs to be the same as the argument's 'self.

In this patch, I handle this case by replacing the argument's 'self with the free region that has already been constructed for 'self in the enclosing scope.

Note that I remove the region param from static trait methods (line 360 of collect.rs). The main motivation for that change is to provide a more sensible error message on something like FromVec::fromVec::<'self>(self.v), which would otherwise confusingly yield the error message "this trait has a lifetime parameter but no lifetime was specified". Moreover, although it could be argued that these methods technically still are region-parameterized, the actual plugged-in region is always the same thing---a free region associated with 'self; I think it's prudent to syntactically prevent anyone from trying to plug anything else in.