Improve couch_btree:chunkify logic

The current chunkify has problems when reduce functions create large values in that it will produce chunks (ie, kp nodes) that contain a single key. In some pathological cases this can create long chains of nodes that never branch.

The old chunkify would also try and create nodes with an even number of bytes in each chunk. Given that we don't re-use chunks it makes more sense to try and pack our chunks as close to the threshold as possible so that we're creating fewer branches in our tree.