Summary sketch

[NelsonVides]

This uses now ddskerl, an implementation of the DDSketch algorithm, instead of quantile_estimator, which is in turn an implementation of Cormode's biased algorithms.

The most important difference is that the biased algorithms aren't mathematically mergeable, hence keeping a per-scheduler summary would never work, and keeping a single global one would never scale (and not mentioning race conditions on inserts as the data-structure is not possible to treat atomically in an ets table). Instead, the DDSketch algorithm is fully mergeable, and the data structure implementation has an ETS backend fully tested for all sorts of race conditions.

As a possibility that is out of the scope of this PR, a custom exporter might provide the Sketch data structure to datadog, so that the quantile merges can be done on the server, and that way the metrics server can aggregate quantiles across many hosts in a meaningful way.

[codecov]

Codecov Report

Attention: Patch coverage is 94.95798% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/metrics/prometheus_quantile_summary.erl	92.77%	6 Missing ⚠️

Files with missing lines	Coverage Δ
src/prometheus_sup.erl	77.41% <ø> (ø)
src/prometheus_time.erl	97.36% <100.00%> (-0.07%)	⬇️
.../eunit/format/prometheus_protobuf_format_tests.erl	100.00% <100.00%> (ø)
test/eunit/format/prometheus_text_format_tests.erl	100.00% <100.00%> (ø)
...eunit/metric/prometheus_quantile_summary_tests.erl	99.09% <100.00%> (+0.55%)	⬆️
src/metrics/prometheus_quantile_summary.erl	94.53% <92.77%> (+2.22%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[lhoguin]

What problem are you trying to solve in this PR?

[NelsonVides]

What problem are you trying to solve in this PR?

Having quantile summaries entirely. The previous ones were known to be broken (see #170, #146, #159). My fix at #170 was just to stop it from breaking, but they were still not correct due to data-loss during race conditions, and surely not performant as they were doing lots of lookup-inserts of big data structures, when this PR does only update_counters.