This one is about a million times simpler, and much more likely to be
correct too.

Instead of trying to match up a tree object against the index, we just
read in the tree object side-by-side into the index, and just walk the
resulting index file. This was what all the read-tree cleanups were
all getting to.