[Java Broker] Stop ServerEncoder on AMQP 0-10 from unnecessarily allocating ByteBuffers

As reported by this email thread the broker can currently run out of direct memory without flow-to-disk ever kicking in.

A defect in SeverEncoder#init() means that we allocate a new 8 kB direct bytebuffer everytime it is called (i.e. on every AMQP 0-10 method or command). This is wasteful and in some cases the GC does not collect these direct byte buffers fast enough leading to the OOM.