Occasional test failures due to Hypothesis' health check

From time to time we get failures in tests running locally or in Travis' builds because the generation of data with Hypothesis is too slow. The canonical failure message is: Data generation is extremely slow: Only produced *N* valid examples in *M* seconds (0 invalid ones and 0 exceeded maximum size). Whether we get this failure or not depends on multiple factors: the machine the tests are running on, the load the processor has and maybe D. Trump's mood or how the dinosaurs actually became extinct.

The type of tests where this happens is very clear: those using compound data. For example, in test_PMap_s1s we give the test one pmap that is produced with a hypothesis composite. The composite needs to produce a (small) number of peaks, which come themselves from other composites and so on. Every composite has been carefully thought so the number of draws is kept small, and still, here we are.

I know we can turn off the health check for any individual test we want, but I want to ask whether we can soften a bit the Health Check globally. I have another case where this happens in #426.

Start with the test_PMap_s1s case and the other compound-data tests described in the issue, then reproduce the intermittent Hypothesis health-check failure locally or in Travis. Review the related discussion in issue #426 and determine what global mitigation is acceptable; done means the affected tests no longer fail intermittently without unnecessarily disabling health checks.