Lots of stub fixes

[o11c]

This fixes a lot of stubs so that they actually typecheck.

The only thing I didn't fix was sqlalchemy, for which I filed #14

Of note:

• Correct the name of python2 __builtin__ so that other code can import it.
• Apply all updates from mypy stubs where they are newer.
• Remove several functions from python2 stubs that only exist in python3.
• Move several stubs to the correct directory (builtin/ vs stdlib/, 3 vs 3.4).
• Use types instead of Any or nonexistent names like builtins.module.
• Fix some incorrect use of Tuple[T] (fixed-length) and also some other sequences.
• Remove a couple of other Any and unparameterized containers that I happened to see while applying other updates.
• Make sure that builtins does not export typing names. There was a comment that implied that imported names were not reexported, but that is not what current tooling actually does, and there were several stubs as well as real-world code that used names like List and Dict without importing them.

What I did not do:

• Verify that all the functions listed in stubs actually exist (I am working on a tool to do this automatically).
• Verify that the types perfectly match (in particular, there are likely missing bytearray and other buffer cases).
• Make any effort to deal with modules that exist in multiple python3 versions but with different contents.
• Reason about platform-specific modules.
• Touch the weird raise FooError code that I don't think belongs in stubs (and which is certainly incomplete, if not inaccurate, if it is intended to enumerate all the possible exceptions).

[o11c]

Note: with the enum commit, I have now tested this against all of: pypy2 (2.6), pypy3 (2.4), python3.2, python3.3, python3.4, and python3.5.

[matthiaskramm]

Thanks for the PR! Your changes look good (a few minor comments above).

Changing builtins.py to __builtin__.py sounds good to me, but it'll need a change in mypy as well. I'd recommend doing that in a separate PR, after python/mypy#884 is merged.

[o11c]

It doesn't need to wait, because a stub-override directory can take precedence.

I have stub-override/2.7/builtins.pyi as a symlink to typeshed/builtin/2.7/__builtin__.pyi, then stub-override/2.7/__builtin__.pyi as from builtins import *.

IMHO the right level of balance is: while typeshed should have consideration for implementations, it shouldn't make itself wrong for them.

[o11c]

I haven't rebased this yet because I'm still investigating a proper fix for the "don't export unless" import as problem.

[o11c]

If I don't finish this in an hour, I probably won't until Monday.

I discovered that mypy does have a remove_imported_names_from_symtable, but it seems ineffective and is wrong anyway.

[matthiaskramm]

No hurry, I won't get around to testing this PR with mypy before Monday, as well.

WRT __builtin__.pyi, let's just put a temporary symlink, builtins.pyi -> __builtin__.pyi into place in typeshed, to not make merging of python/mypy#884 more complicated than it needs to be.

[o11c]

A symlink is not enough, since that will result in duplicate definitions if anybody explicitly imports it.

I'll use the (working) glob import though.

[o11c]

Switching from my versions of import as fix to Jukka's was much more painful than I expected.

I have addressed all of the review comments now except for the typing.IO case, since everybody is forced to be wrong there, and the way I wrote it will make an eventual fix easier.

There's actually already a bug: python/typing#112

[o11c]

Note, the hang you experienced is a result of creating a subprocess.Popen with stdout=PIPE but not reading from it until the subprocess dies.

[JukkaL]

FYI, mypy should now deal with import as correctly in stubs.