Previously we were optimizing for cpu time, so we where building
gst-build once and then exporting that to be used by the test jobs.
However this meant that we where uploading 200mb (previously 600mb)
zipped of artifacts and then re-downloading them for each test job.
This caused big costs in terms of cloud egress since the runners
aren't hosted on the same cloud as the storage/artifacts instance.

Instead we are going to be rebuilding gst-build for each test
job from now, it also doesn't take more time than the network
i/o would of downloading the artifacts, so the impact of rebuilding
shouldn't be noticebly.

We are also using pinned git refs the modules we rebuild from
the manifest, so the binaries should be reproducible for the most
part (minus things like .pyc files).

Close #68

Part-of: <!280>