Certain cases, like diffing an empty buffer with a big buffer, were
very slow, now this should get better as we will directly detect the
matching eol at the end of both buffers, end then immediatly detect
we need to add the rest of the big buffer.
We still are too slow on some general diff when there is a lot of
differences.
We can index native arrays negatively, so just setup V1 and V2 to
point in the middle of the work arrays and remove the need for
creating MirroredArray.
