Pluggable buffer for network actors

Cluster singleton and cluster sharding both uses an internal buffer (and not the mailbox) for when the remote actor is not available. It would be nice if that buffer was pluggable (as the mailbox were).

Use-cases:

• If you have a retry mechanism, you may want to throw away new messages when the buffer is full. That way the retry will not effect message ordering too much.
• If you send updates for some entries, and are only interested in the latest update for each entry, you could implement a buffer that throws away all out-of-date information and only tries to send the latest info.
• Maybe you don't want to change the behavior when the buffer is full, you just want to notify someone (logging, back-pressure etc).

What do you think?