exact pixel boundaries when we walk the inverse matrix (e.g. 2/3, 1+1/3, 2, 2+2/3, 3+1/3, 4, ...).
These are not stable, since any variance in the inverse matrix (even in the lowest bit) can cause some other
part of the ninepatch to start a hair to the left, resulting in misaligning every 3 pixels.
The fix changes the matrix' phase enough to keep all of the stretched ninepatches in phase.
b5c17a64ed