Previously if HTTP/1 proxy is used for backend connection, we read all
incoming bytes from proxy including response body, which may be part
of HTTP/2 protocol.  While investigating this issue, we found that
http_parser_execute() returns 1-less length when we call
http_parser_pause() inside on_headers_complete callback.  To
workaround this, we increment the return value by 1.  This commit also
fixes possible segmentation fault error, which could be caused by the
lack of stopping libev watcher in disconnect().

Previously we did not handle the situation where RST_STREAM is
submitted against a stream while requet HEADERS which opens that
stream is still in queue.  Due to max concurrent streams limit,
RST_STREAM is sent first, and then request HEADERS, which effectively
voids RST_STREAM.

In this commit, we checks RST_STREAM against currently pending request
HEADERS in queue and if stream ID matches, we mark that HEADERS as
canceled and RST_STREAM is not sent in this case.  The library will
call on_frame_not_sent_callback for the canceled HEADERS with error
code from RST_STREAM.

Previously we use 16K - 9 bytes (frame header) as frame payload size
so that whole frame fits in 1 TLS record size (16K).  But it turns out
that in proxy use case, we will receive 16K payload from backend and
we have to split it into 2 odd looking frames (16K - 9 and 9), and
latter is highly inefficient.  To avoid this situation, we decided to
use min frame payload size to 16K.  Since we operates on TLS as stream
of data, we are not so much restricted in its record size.

We missed wb.reset(), which makes write buffer's capacity becomes 0
and communication stalls eventually.