Use a sync.pool for memd.conn readers

At present if we have a large number of connections per node (e.g. a number of agents each with a kv pool size of say 8) and something goes wrong and we end up in a reconnect loop for a couple of minutes then we can up blowing up our memory footprint as we create a 20MB buffer per connection.

We should move to using a pool so that we reuse the buffers.