If e.g. the right edge of the leftmost rectangle is near MIN_INT, and
the left edge of the rightmost rectangle is large then subtracting these
can lead to an integer overflow, making the resultant "width" falsely
positive, thus returning a very wide result instead of the expected
no-intersection result.
We avoid the overflow by not doing the subtraction unless we know the
result will be positive. There are still risks for overflow if x + width
or y + width is larger than MAXINT, but we won't ever overflow for valid
rects now.