Add an overflow check in the Iter::next() impl for Range<_> to help with vectorization.
This helps with vectorization in some cases, such as (0..u16::MAX).collect::<Vec<u16>>(),
as LLVM is able to change the loop condition to use equality instead of less than and should help with #43124. (See also my [last comment](https://github.com/rust-lang/rust/issues/43124#issuecomment-319098625) there.) This PR makes collect on ranges of u16, i16, i8, and u8 **significantly** faster (at least on x86-64 and i686), and pretty close, though not quite equivalent to a [manual unsafe implementation](https://is.gd/nkoecB). 32 ( and 64-bit values on x86-64) bit values were already vectorized without this change, and they still are. This PR doesn't seem to help with 64-bit values on i686, as they still don't vectorize well compared to doing a manual loop.
I'm a bit unsure if this was the best way of implementing this, I tried to do it with as little changes as possible and avoided changing the step trait and the behavior in RangeFrom (I'll leave that for others like #43127 to discuss wider changes to the trait). I tried simply changing the comparison to `self.start != self.end` though that made the compiler segfault when compiling stage0, so I went with this method instead for now.
As for `next_back()`, reverse ranges seem to optimise properly already.
0f9317d37e