Add rounding epsilon for NEAREST filter after transformation, not before.

Add rounding epsilon for NEAREST filter after transformation, not before.

For nearest filtering, the transformed location must be rounded to the
nearest sample. For locations that precisely in the middle between two
samples, we round down by adding -epsilon to the coordinates.

Before, we would do this on untransformed coordinates. This patch
changes it to happen after transformation. It also clarifies the
difference between sample locations and rounding.

Reported by Robert O'Callahan.