Don't align the area we write to the target to the block size. If the format matches, the writes don't need
to be block aligned. We still read the whole thing in, because that's the granularity that ReadTexture
operates at, but discard those pixels when updating the framebuffer. Onimusha 2 does this dance where it
uploads the left 4 pixels to the middle of the image, then moves it to the left, and because we process
the move in hardware, local memory never gets updated, and thus is stale.