This changes makes sure we never write to *writeData in the ServeHTTP
goroutine until the serve goroutine is done with it.

Also, it makes sure we don't transition the stream to the closed state
on the final DATA frame concurrently with the write.

To fix both, the writeFrameAsync goroutine no longer replies directly back
to the ServeHTTP goroutine with the write result. It's now passed to
the serve goroutine instead, which looks at the frameWriteMsg to
decide how to advance the state machine, then signals the ServeHTTP
goroutine with the result, and then advances the state machine.

Because advancing the state machine could transition it to closed,
which the ServeHTTP goroutine might also be selecting on, make the
ServeHTTP goroutine prefer its frameWriteMsg response channel for errors
over the stream closure in its select.

Various code simplifications and robustness in the process.

Tests now pass reliably even with high -count values, -race on/off,
etc. I've been unable to make h2load be unhappy now either.

Thanks to Tatsuhiro Tsujikawa (Github user @tatsuhiro-t) for the bug
report and debugging clues.

Fixes golang/go#12998

Change-Id: I441c4c9ca928eaba89fd4728d213019606edd899
Reviewed-on: https://go-review.googlesource.com/16063
Reviewed-by: Andrew Gerrand <adg@golang.org>