Expand coverage of stability integration tests

[fjetter]

Since the stability integration tests are now established, we should focus on extending the capabilities and coverage of this test suite to maximize its value.
Newly added tests should focus broadly on capabilities dask offers and not focus on specific implementation details. Ideally tests focus on stressing specific areas of the system in a way that is typically hard to reproduce on a unit test level.

• Integration tests: Adaptive scaling #134
• Integration tests: Graceful worker restart #135
• Integration test: common physical science workload #174
• Integration test that focus on AMM informing our decision to toggle this on by default #140
• Integration tests: spill/unspill #136
• Integration tests: GIL #137

Won't fix

• Almost embarrassingly parallel workload piles up on a single worker #141