When you have a cluster in Infinispan, configured as Distributed and with a single and shared cache loader by all members, the preload option should work making each node loading the entries that correspond to it, not all of them, which is the current behaviour.
Also, with this configuration, the "state transfer" could be configured to make each node retrieve the "new entries" that should keep after a topology change directly from the cache store, instead of the transfer of entries through the network.