limit check insertion via loop forests

[Stephen Weeks]

> But if there is a non-allocating loop followed
> by some allocations, we want to avoid moving the limit check into the
> loop, where we would expect it to be executed more often.

Although, if there is an allocating loop followed by some allocations,
I would argue that we should move the limit check backward and
coalesce it into the loop limit check.  My proposal doesn't do this,
so there is still room for improvement.