Ensure HTTP/2 connections are gracefully closed on async cancellation #756

• Initially raised as discussion Broken HTTP2 cancellations tests #755

The issue is that some HTTP/2 connections can fail and not be cleaned up, resulting in our connection being in a broken state.
When the asynchronous semaphore acquisition is cancelled, our handle_request method propagates the exception and never calls clean up.

The first step, as always, will be to write a failing test, but in this case that will be difficult because we need to mock the AsyncSemaphore object, more specifically, we want the acquisition method to be extremely slow.

Let's use monkeypatching to write a failing test with no complex changes, and then we can change that failing test if we want.