Declare dependency on lxml

[rowillia]

Mypy depends on lxml for reports but doesn't declare it as a dependency
so it doesn't get installed alongside of mypy. Also, mypy no longer
works with latest release of lxml so I had to pin to the previous release.

[JukkaL]

Thanks for the PR! Have you tested that this works in Windows? I've heard people complain about installing lxml in Windows.

In the longer term I'd rather get rid of the lxml dependency (see #909).

[rowillia]

@JukkaL I don't have a windows box :(.

If we don't want to support depending on lxml, should we disable the reports that need it if it isn't installed?

[JukkaL]

Yeah, detecting if lxml can be imported and falling back to more basic behavior would be totally fine.

I have a Windows VM and I can try this later when I have time, but I have a bunch other stuff I need to finish first.

[gvanrossum]

@rowillia in #1120 I'm fixing the problem reported in #1002. Can you review that? (I'm not so clear on XML, but I got it to install and work. :-) Once that's in, you can redo this one without the pinning.

[gvanrossum]

This is still awaiting a test for Windows. Honestly the report generation code is so flaky that I don't think we should merge this at all. However, I fixed the lxml issue, so that it now works with lxml 3.5, so at least the pinning is no longer necessary.

(FWIW I also added some more pep8 exceptions so the flake8 pinning is no longer necessary either. And somehow the Travis CI setup wasn't using test-requirements.txt anyways.)

[gvanrossum]

I'm closing this -- we should do this differently anyway, by removing our lxml use.