Improve performance of json encoding in go_json

One of the things that stifles N1QL throughput is results son encoding: to put it in perspective, in a single node test environment, returning a single marshalled field, after encoding has been measured at about 350KB/sec, vs returning one large non marshalled document at a respectable 230MB/sec.
While the process of unmarshalling and marshalling will never be able to match raw document performance, it should be doing much better than what it currently does.

This MB keeps track of the marshalling side of things.