Integration tests are very slow and test bench build is not cached properly locally

[ml-evs]

PR #160 comes with a fairly significant performance cost on the integration tests. These are not run by default in the local test suite (can be forced with env var CI=1), but this should probably be revisited soon. There is discussion in the comments of #160, but it seems the main slowdown is the fact that docker-py does not seem to be make good use of the existing cache, perhaps due to using some features not fully supported (docker-py cannot use buildkit).

We can switch over to something that maps more closely to the Docker CLI, which should improve performance. More generally, the integration tests can be a bit slow as the containers are simulating real queues that have their own polling times. At the very least, we can parallelize the integration tests over Slurm/SGE, though this will be a bit fiddly with the current setup. I've added the pytest-durations plugin to track these slow steps.