Refactor the collectors/outputs

[na--]

It's becoming clear that the current Collector interface isn't very fit for purpose. It is a poor fit for a lot of the outputs ([1], [2], [3.1], [3.2], [4], [5]), it leads to some code duplication and makes correctly writing new thread-safe outputs relatively difficult, and it also doesn't offer enough functionality for more complicated outputs (#1074). The same thing even has multiple names - what we call Collectors in the code live in the stats folder and are presented as "outputs" to users...

Additionally, the thresholds and the end-of-test summary are basically somewhat special-purpose outputs... They ingest metrics and do some calculations on them - the summary just spews the aggregated results at the end of the test, while thresholds do some calculations and potentially abort the test execution... They are enabled by default, but can be disabled by users. I'm not sure if it makes sense to refactor them into outputs for just a marginal improvement of code quality, but it should at least be investigated...

I think that when we're investigating the telegraf integration (#1064), we should also investigate their output pipeline architecture. Since their main job is dealing with metrics, there are probably things we can learn from them...

Other somewhat related issues: #917, #570, #587, #1001, #743, #572, #355, #351

[na--]

For completeness' sake, other reasons to refactor the current interface and architecture of outputss: #1606 and #1423

[na--]

Another reason to refactor the output interface is that the k6 outputs currently don't whether the test run finished successfully, and if not, what was the error. This is preventing us from being able to send the error (e.g. an exception in setup() or teardown()) to outputs that will be able to understand and record it (e.g. the cloud).