The benchmarks were not working because they couldn't access Git
information. While this is fair, they didn't actually fail, they kept
going, so I have improved the script to fail hard, and fixed the bug by
setting `RELEASE_VERSION` so the build script doesn't bother trying to
read Git.
V3_GIT_ORIGIN_REV_ID: 9573ea32371d7a4d7b99c87017a8d5d77815832c
This splits out a `debug.Dockerfile` which makes use of out-of-band
caching to speed up builds drastically, at the expense of
reproducibility.
It is used to run tests and auxiliary test services (i.e. the custom
connector).
The new `debug.Dockerfile` marks the Cargo dependency and build caches
as Docker caches, which means they are shared between builds. This is
probably fine for local work and testing. The `Dockerfile` continues to
not use a cache like this, to guarantee that it is not polluted by extra
information, at the expense of build speed.
In addition, we build a `nextest` archive ahead of time to avoid
building tests when attempting to run them.
On my machine, a re-run of `just test` now takes seconds.
I have also sped up the `postgres` container start time by creating a
database called "finished" last, and then waiting for that to show up.
V3_GIT_ORIGIN_REV_ID: 7ef0548361987175b68a0cad44c8f2295110a1fb
