Add Travis job with 3.6 nightly

[jenshnielsen]

As suggested in #5009

This job is allowed to fail and I expect it to fail due to the removal of the deprecated inspect.getargspec
I have also moved the --pre flag to this job from the 3.5rc2 one

[tacaswell]

I think we need the pre flag to get numpy to install?

[jenshnielsen]

@tacaswell You are right. I am a bit unhappy that the pre flag was repossessed for that purpose. I have added it back for now but I would like to change the wheelhouse behaviour in a future PR

[jenshnielsen]

I fixed the majority of the 3.6 test failures by replacing getargspec with getfullargspec This is not long term the correct solution since getfullargspec is deprecated too, but rewriting the code to use signature is not something that I am comfortable with doing until after 1.5. I think this should be safe for the 1.5 since getfullargspec is basically just an extension of getargspec I think it would be nice to get this into 1.5 (and 2.0)

The remaining failures are due to nose depending on getargspec too

Edit: Clarified that the issue with signature is purely timing with respect to 1.5

[tacaswell]

My only concern is that this is going to conflict with the changes to boilerplate in #4829

[jenshnielsen]

Makes sense. I suggest waiting with this one until #4829 is merged. Rebasing from this side should be much easier.

[jenshnielsen]

Rebased after #4829 was merged, It's probably to late now but it would be nice to get in before 1.5 since matplotlib basically don't work without in Python 3.6. Most of the tests pass now but nose still uses getargspec so a few of the tests fail

[tacaswell]

This is a bugfix.

[tacaswell]

Drat, I was fooled by the green-check on this page, there are still a few instances of getargspec in the code base so we fail on 3.6

[tacaswell]

And I read the failure too fast and now see that these are the failures due to nose that @jenshnielsen mentioned above, sorry for the noise.