Forwarding for non-transactional write commands should be asynchronous

A write command keeps a lock on the originator while doing the RPC on the actual owners, so forwarding the command back to the originator synchronously leads to a deadlock.

Keeping track of the real originator for forwarded commands could allow us to avoid the deadlock, but it would be too complicated. OTOH, non-tx caches don't guarantee consistency, so we could just do the forwarding asynchronously.