Fix issue #231

[lohedges]

This PR closes #231.

_mix_replicas returns the state each replica's configuration moves to, but the mixing code treated it as the replica each state is seeded from, so it applied the inverse permutation. That's harmless for pairwise swaps but wrong for chained swaps around three or more replicas, which are common when energies are close and many swaps are accepted. Configurations were sent to the wrong lambda windows.

• _mix_replicas now returns the inverse (np.argsort), matching how the states are used.
• The transition matrix now records each configuration's move. It previously paired states from two consecutive cycles.
• Legacy checkpoints (pre-slots) are inverted on restore.
• Added a test that runs mixing and state application together for a 3-cycle.