Fix problems with multigets and replication

Fix problems with multigets and replication

The current code could go out of sync and send multiple "NOOP" to the same
server depending on the replica number to get, and the fetch-code isn't
capable of handling that. The memcached_mget_execute code also complcates this
because it could process some of the messages while receiving them so we
cannot reset the get sequence if we are able to send out the commands but
not the NOOP... To create a sane patch for this I disabled the buffered mode.
It will increase the number of packets sent to the server, but at least we
have a well defined behavior.