Output json optimizations

[mstoykov]

Mostly provoked by this community forum post.

Moving to using easyjson gave some improvement, but then I noticed we are spending a bunch of time in compression so decided to try out a different gzip implementation.

Combined improvement is twice better on both a CPU and allocations

name            old time/op    new time/op    delta
FlushMetrics-8    37.2ms ± 6%    18.2ms ± 6%  -50.99%  (p=0.000 n=10+9)

name            old alloc/op   new alloc/op   delta
FlushMetrics-8    2.44MB ± 1%    2.52MB ± 0%   +3.28%  (p=0.000 n=10+10)

name            old allocs/op  new allocs/op  delta
FlushMetrics-8     47.3k ± 1%     20.1k ± 0%  -57.59%  (p=0.000 n=10+10)

Breaking change: the tag keys are now not sorted in the output. This is due to no longer using the stdlib marshalling which does this.

Changelog entry under "new features" (I guess):

The JSON output was optimized and now should around 2x as more performant at outputting metrics. 

This means that it either will be able to export twice as many metrics and/or to use half the resources to do the same amount of metrics.

As a side effect there is a slight breaking change in which the `tags` field is no longer sorted, while it was previously. 

[mstoykov]

Some of those can likely be used in the output/cloud as well as they don't seem that complicated to apply

using jwriter is like 10-15% faster and kaluspost gzip (which is the majority of the additional lines, unfortunately) seems to give like 20% better performance so that seem also quite nice.

[na--]

LGTM, though please add a PR description with a short description and a benchmark comparison for the cumulative speedup of all commits

[mstoykov]

@na-- updated

[mstoykov]

Note: if we change the time to be int64 instead of time.Time we can get even better performance
Compared to the optimized version here:

name            old time/op    new time/op    delta
FlushMetrics-8    18.2ms ± 6%    11.3ms ±12%  -37.96%  (p=0.000 n=9+10)

name            old alloc/op   new alloc/op   delta
FlushMetrics-8    2.52MB ± 0%    1.85MB ± 0%  -26.64%  (p=0.000 n=10+9)

name            old allocs/op  new allocs/op  delta
FlushMetrics-8     20.1k ± 0%      7.6k ± 0%  -62.31%  (p=0.000 n=10+10)

Compared to the original:

name            old time/op    new time/op    delta
FlushMetrics-8    37.2ms ± 6%    11.3ms ±12%  -69.59%  (p=0.000 n=10+10)

name            old alloc/op   new alloc/op   delta
FlushMetrics-8    2.44MB ± 1%    1.85MB ± 0%  -24.23%  (p=0.000 n=10+9)

name            old allocs/op  new allocs/op  delta
FlushMetrics-8     47.3k ± 1%      7.6k ± 0%  -84.01%  (p=0.000 n=10+10)

Unfortunately I can't figure out a not terrible way to support both so we can be backwards compatible :(

Also, of note is that I remembered there is a slight breaking change - the tag keys are now not sorted in the JSON output.