I changed memg.py to use sockfile.write()/.flush() instead of socket.sendall().

This makes the memg.py server > x100 faster. It outperforms a gevent-based implementation by 10%.

See https://github.com/codeape2/Key-Value-Polyglot/commit/cbc53a...

EDIT: It does not outperform the gevent-based implementation. More performance testing indicates that gevent is around 2x faster. But it outperforms the original version by an order of magnitude.