The unit tests were failing due to being unable to connect to the
pubsub-emulator service. The problem seems to be that the service
container was taking an unreasonably large amount of time to start, and
the tests would attempt to connect to it before it'd begun listening.

Configuring a `HEALTHCHECK_TCP_PORT` for the service containers results
in the main job steps not starting until the service containers have
started listening. This allows the tests to pass, but currently the
tests take about 4x longer to run than when they used to run on the
on-prem runners rather than the GKE-hosted Kubernetes executor.

Although the pubsub-emulator service takes a minute or two to start with
the tests passing, this is still less time than the tests previously ran
for while retrying connection attempts to the service. So I'm not
entirely confident that something else is going on too. It could be that
resource contention was even higher previously when the test had begun
running, leading to the service never being able to start.

Fix failing precommit and python:tox CI jobs

See merge request !75