Thought: do we want to just `impl Drop for TestHarness`? This might make the panicking situation tricky, whereas explicitly calling `harness.finish()`, while possible to forget, makes handling panics just work.

Ensures shutdown and cleanup so that we can't forget
This is important during benchmarking; Criterion creates thousands of servers, and if we don't shut them down and clean them up, we'll exhaust the system's process pool