tests should fail for good reasons only

[mdeff]

Fixes the following test error that happened far too often.

[coveralls]

Coverage increased (+0.002%) to 88.594% when pulling 43dc6d5 on deterministic-reduction-docstring into f6b5a1f on master.

[mdeff]

And that's the key retrieval issue that often pops up as well.

[mdeff]

Another kind of false test failures is the verification of hyperlinks. But that's pretty useful as hyperlinks tend to break now and then. We could run it without enforcing it to succeed, but I'm not sure whether we'll pay attention to it if it never fails.

http://tulip.labri.fr fails especially often.

[mdeff]

Newer sphinx will retry failed hyperlinks (to handle rate-limits) in linkcheck sphinx-doc/sphinx@6b90a63. That should also help.

[nperraud]

All is good.

Good improvement to the function graph_sparsify

Thanks