This commit synchronizes the in-memory cache with the
on-disk state to ensure the waiting proof store is
externally consistent. Currently, there are scenarios
where the in-memory state is updated, and not reverted
if the write fails. The general fix is to wait to apply
modifications until the write succeeds, and use a
read/write lock to synchronize access with db operations.
In this commit, we modify the waiting proof slightly to acept dupliacte
waiting proofs, rather than reject them. Otherwise, it's possible that
the remote node first sends us their half of the waiting proof (before
we do), we write that to disk, then upon restart, we'll try to add it
again, but be rejected by the system.
Fixes#1315.
