This gives what is to me smoother and more predictable snapping motion with
respect to the current mouse position.

For example-- imagine the grid cell size is 10 and the selected object under
the mouse already has its top left corner aligned with vertical grid line #2.

If I begin dragging that object to the left, within a single pixel of motion
it will snap to vertical grid line #1 which is 9 pixels away from the
object's current position.

With this commit, the user must drag the object cellsize / 2 pixels before it
snaps to a new position.

To me this seems smoother as it rounds to the nearest grid line rather than
to the grid line with the smallest coordinate value.