Smoothing regulariser fails in ML engine test under certain circumstances

Noticed some funny behaviour with the ML engine integration test when using the smoothing regularizer. It seems to fail in github actions with the core dependencies but not with the full dependencies. In particular, the extra `matplotlib` dependency seems to "fix" the issue. Unclear what exactly is happening here. Have disabled the test for now.