tdb_cons_finalize() behaves badly with read IO if data doesn't fit in RAM

[tuulos]

It appears that random access to the big mmapped items buffer, e.g. in edge_encode_items causes a very suboptimal read IO pattern if the mmapped area doesn't fit in memory, making encode_trails extremely slow under some conditions.