Currently, mixing futures and streaming_callback does not guarantee
ordering (so the returned future could become ready before the last
streaming_callback fires). Specifically, this happens with
test_streaming_read_until_close_after_close if that test is modified
to not use the callback argument. This would cause problems when we
remove the callback argument if we left streaming_callback in place.

These problems are solvable, but probably not worth it since
partial=True is an alternative (and used internally instead of
streaming_callback).

This requires a lot of mechanical changes throughout the test suite.
tornado.websocket is currently still using the deprecated interfaces
and suppressing the warning.

This is mainly preparation for HTTP/2: request.write was designed to
mix headers and bodies in a way that is no longer supported.

This method is not actually redundant with
HTTPMessageDelegate.on_connection_close (maybe there is a useful
refactoring that would make it so, but for now it's still useful)

test: Reinstate a test that was mistakenly left disabled

This reportedly works around https://github.com/pypa/pip/issues/5240

This test required some refactoring to construct invalid headers that
we no longer allow handlers to return directly.

docs: Remove repeated lines.

Added RequestHandler.on_finish() to FallbackHandler.preapre()

Add flow control to websocket client

It removes repeated lines in
docs for release 5.0.0

reduces likelihood of alignment with many similar timers

httpclient: Deprecate fetch callback, narrow `raise_error`

This code is currently not exercised in CI.

This avoids collisions in code that must deal with both
httpclient.HTTPError and web.HTTPError.

Give it the same deprecation behavior as AsyncHTTPClient.fetch

ioloop: Fix PeriodicCallback when system clock moves backwards

If the clock moves backwards a little bit, we could run a
PeriodicCallback twice in a row prior to this change.

Due to implementation differences between the Tornado and asyncio
event loops, this is most common when using the asyncio event loop on
windows (where in some environments differences between time.time()
and time.monotonic() can result in behavior that looks like small
backwards time jumps on every iteration), although it can in principle
occur in any environment.

Fixes #2333

This permits it to be tested separately from the (now python 2.7-only)
IOLoop configuration.

asyncio: Fix a leak when event loops are created and destroyed

Eliminate a confusing error message when this occurs.

Fixes #2325

The WeakKeyDictionary in IOLoop wasn't doing its job because of
reference cycles. This was easiest to see with the synchronous
HTTPClient.

Fixes #2321

testing: Don't close the event loop if we didn't create it

This improves compatibility with asyncio test frameworks like
pytest-asyncio.

Fixes #2324

gen: Deprecate legacy interfaces

This temporarily duplicates the tests, but the versions using the
deprecated interfaces will be gone in 6.0