Reduce amount of objects allocated by the codec

[2opremio]

Closes #972 . Improves #812 and #854

Results in the probe generating ~20% less objects. See ugorji/go#143

[paulbellamy]

As discussed in slack, the decoding thing will only affect gzip writers, which we might be able to solve with a bufio.Writer wrapping the gzip writer. the stringView one is a bit trickier. :|

[2opremio]

Unfortunately ugorji/go#143 shows the biggest benefit when encoding. So ... should I revert it?

[2opremio]

After thinking about it a bit more, we already obtain much higher performance from codecgen because we run it with -u, which generates code applying the same kind of tricks. So I don't think -tags unsafe makes the situation worse (and I certainly wouldn't consider using codecgen without -u for the reasons explained).