[QA] CI response time investigation and improvements

[Legoktm]

I think the most obvious improvement would be to stop spinning up an
entire VM for individual jobs. I think at the time nodepool was a good
idea, but the benefits we were supposed to get (e.g. getting rid of the
whitelist) haven't happened, and it's causing problems on delaying CI
and preventing us from adding new jobs (e.g. PHP 7) that are pretty
important IMO.

I understand that there was work on a docker-based solution that should
be faster and not as expensive, has any other progress been made on it?

-- Legoktm