Implement XML-based reports

[o11c]

This implements a new reporting framework and uses it to implement XML reports using lxml. The old HTML report is still available using --old-html-report <dir>.

The new reports implement --xml-report <dir> and --xslt-html-report <dir>, and makes --html-report <dir> an alias for the latter. It is anticipated that --xslt-txt-report <dir> could be written, and also that a 3rd party might write their own XSLT file that combines mypy's report with coverage's report in a single HTML document.

The XML report embeds a reference to the XSLT document, so it can be viewed directly in most browsers. However, WebKit-based browsers (such as Google Chrome) have a bug that prevents this from working on file: URLs. For this and text-based browsers, the XSLT-based html report internally performs the transforms, and also rewrites the extension of the links from .xml to .html. An improved external CSS stlylesheet is used regardless of who does the XSLT step.

A single XML document is constructed in memory (potentially useful directly for unit tests, though this patch doesn't include any) if any of the new transformations are requested, and written or translated during output. If none of the new reports are requested, lxml does not need to be installed.

A very minimal stub for lxml.etree is included at top-level. It should probably go somewhere else, but there is no precedent for a non-standard-library stub.

While writing this, I noticed that it would be useful to use python3.4's enum (which is backported as enum34). This likely requires special support because it uses a metaclass to do weird stuff.

This PR includes the commits from #711 since it should be merged ASAP, and also some fixes to stubgen that could be refactored into a separate PR since I didn't end up using them because they ended up crashing mypy once I started specifying some of the arguments it couldn't detect.

[spkersten]

I think that your str annotations could (should?) be AnyStr.

[o11c]

Changed str to AnyStr. Didn't test every function, just one, but I assume lxml is consistent.

That reminds me, bytes(foo) should work when type(Foo) has a __bytes__ method, I currently do an unnecessary encode/decode cycle because of its lack.

Also implemented --txt-report (as proof-of-concept; it only outputs a summary, not per-file info) and refactored the argument logic a bit.

[o11c]

travis says: lxml/etree.pyi:17: error: Incompatible types in assignment (expression has type "str", variable has type "bytes")

This is a bug in other parts of mypy, since this code should be okay: def write(method: AnyStr = 'xml): pass

[o11c]

@spkersten Okay, I've figured out that we don't wany AnyStr here, since that's a TypeVar. Instead we want a Union[str, bytes] which is completely different.

[o11c]

Ah, I didn't realize there were two copies of typing that had to be changed. That's really bad.

[JukkaL]

I'll look at this now that the test driver PR was merged.

[JukkaL]

I rebased and pushed the PR, after having tweaked the CSS a little further.

I find XSLT painful to use and most contributors won't be familiar with it (see https://www.google.com/trends/explore#q=xslt), so I'd like to get rid of it eventually, but we can live with it for a while (I created an issue for it: #909, but it's not really urgent).