Implement _bisect_recursive, which uses multiple bisect calls to
handle renames and finding entries in subdirs.
As is, this could be hooked into paths2ids() if the dirstate has not been loaded yet.
However, it doesn't quite provide enough, since the parsed entries would probably not
be saved. Further, the multiple bisect calls are less efficient then they could be,
because they do not remember the last bisect call.
We should explore switching to a caching structure, which maintains all records that
have been processed, in a structure that can be in-memory searched before going back
to disk.