Backwards adaptation doesn't really make sense in this context,
because ext-tile means the decoder might not decode every tile in the
frame, in which case it can't work out how to adapt probabilities.

It turns out that the tests in the tree were passing when
SIMPLE_BWD_ADAPT was disabled because (presumably) all of them
happened to decode every tile, so backward adaptation worked as you'd
expect. With SIMPLE_BWD_ADAPT, there's some code in get_tile_buffers
to figure out which is the largest tile. This wasn't mirrored in
get_ls_tile_buffers, which caused the encoder and decoder to get out
of sync because they ended up with different CDFs. You also can't
easily fix things in get_ls_tile_buffers because that code is careful
not to calculate the size of tiles it doesn't need.

BUG=aomedia:917

Change-Id: Ia926692f86ca1466252108e09da3de590d45f048