Cancelled all but windows build.

normally I would prefer the numpy versions to be explicit, but since this is a temporarily workaround, <1.18 should be fine.

apparently the < is not OK, so I go ahead and make the changes to get this in asap.

(all, but the windows build is cancelled, merge away once that's passing)

This CI issue will actually be fixed when #9726 is merged (since Numpy won't be installed with conda anymore), so can we hold off maybe so as not to create conflicts? I think @eteq was planning to have one last look at that PR before merging.

Just to be clear I'd still advocate keeping this open though in case #9726 gets delayed beyond a couple of days.

@astrofrog - I would rather merge this away, normalizing having failing CI sets a bad precedent (it's already too bad that we had expected failures for a couple of weeks now with the other jobs), and this conflict would be a rather trivial one for #9726 to resolve (besides this may need to be backported, depending on how long it takes to get the upstream fix).

Ok sounds good - let's do it!

Ok sounds good - let's do it!

Thanks @astrofrog!

Thanks all, I think this is an excellent solution, especially as the comment in .travis.yml makes very clear what we're doing.