Fixes #2283

docs: Updates for sphinx 1.7

This introduced a new inherit-docstring option that we don't want
(it currently breaks our docs build) and got stricter about `:arg`
types.

curl_httpclient: Only call native_str on debug message when needed

Some debug messages are binary blobs that should not be decoded.

Fixes #2275

Fix/etag304

Fix typo in URLSpec.__init__ docstring

StaticFileHandler would return a 304 despite the client sending
a non-matching Etag header when If-Modified-Since matches

fix gen.multi with gen.moment

is_future() only accepts 'real' futures, but NullFutures work here, too.

fixes assertion in gen.multi when calling convert_yielded with more than one gen.moment or None

convert_yielded can be called with None or [None].
When it is called with [None], gen.multi gets [gen.moment],
which fails assert(is_future) since it is no longer a real Future

Update release notes and set version to 5.0b1

test: Adjust unittest for tornado_http2 compatibility

Some recently-introduced tests are failing under tornado_http2, so fix
them.

ioloop: current() and friends delegate to asyncio

It is no longer possible to provide this method with reasonable
semantics.

Instead of a redundant IOLoop._current thread-local, pass through
directly to asyncio and maintain a one-to-one mapping of asyncio loops
to IOLoops. This brings us a bit closer to the asyncio-only future.

Misc small fixes for 5.0

This is currently failing with setuptools errors.

The asyncio coroutine runner cannot be made to use backports_abc.

Now that futures support cancellation, it is unsafe to return the same
Future object to multiple callers.

Fixes #2189

This reverts commit e0c489fd.

That commit was a workaround for old versions of the coverage package.

Fixes #1895

This is additional verification that the asyncio integration is
working as intended.

Fixes #1684

iostream: Add read_into

Tests come from Antoine Pitrou's #2193

Fixes #2176

This has the same memory-allocation behavior as before, but it moves
the buffer out of the recv() call to python code.

iostream: Use file objects instead of raw descriptors in PipeIOStream

These tests are flaky with the new "messages to stderr are errors"
check.

Python 3 logs warnings in destructors if objects responsible for file
descriptors are not explicitly closed. These warnings were previously
being ignored by our test suite unless a human was looking at the log
output, but several recent PRs have introduced these issues. This
change ensures we catch them (and fixes the most recent one in
process_test. The leak has always been there, but the previous commit
caused it to be logged).

This makes the implementation of PipeIOStream more compatible with
socket streams, and allows for more of the test suite to be used with
it.

Extracted from PR #2193

ioloop,gen,asyncio: Improvements to IOLoop GC

Correctly document that the result of Condition.wait() is
not to raise a TimeoutError, but rather False.

Any asyncio event loop that was created but neither closed nor set as
current for some thread will log a warning at the end of the test (in
a way that doesn't cause the test to fail). Fix a bug in
make_current() (in which multiple calls would clobber
self.old_asyncio) and use it more consistently when "current" status
changes.