Unbreak async_io test with hyper 0.14.27+

[acfoltzer]

As noted in #285, a change in hyper's behavior with empty chunked encoded bodies is observable via the assertions in the async_io test. We are pinning to 0.14.26 for now, but obviously would like that to be temporary.

[itsrainy]

Even on the current hyper version this test seems to be flaky on mac, possibly also on linux. On my mac, I ran it 1000 times consecutively and it failed 6 of those times. I haven't observed it failing on my linux machine yet. Anecdotally seems to be more prone to failure in CI (I noticed it failed on #294 earlier today)