Since we do not call on_data_chunk_recv_callback for ignored DATA
chunk, if nghttp2_option_set_no_auto_connection_window_update is used,
application may not have a chance to send connection WINDOW_UPDATE.
To fix this, we accumulate those received bytes, and if it exceeds
certain number, we automatically send connection-level WINDOW_UPDATE.

This commits changes the upper bound of one header field size (the sum
of the length of name and value) to 64KiB by default.  We may add an
option to change this upper bound in the future.

For now, if request has request body, we'll issue RST_STREAM to inform
the peer to stop sending body.  RST_STREAM may be sent before error
page header or data, so peer may receive RST_STREAM only.

The number of failed stream is counted multiple times if several
connection attempts are made.

Previously we do not specify the number of requests each client has to
issue.  The each client corresponds to 1 TCP connection.  If
connection was not accepted by server or not TLS handshake is not
done, we effectively don't use that connection and the requests
supposed to be issued for those connections are done via other
established connections.  If this occurs, servers which do not accept
all connections may gain good benchmark results since they don't have
to pay extra cost to handle all connections (e.g., SSL/TLS handshake).
This change explicitly set the number of requests each client has to
issue so that servers cannot *cheat*.

Previously in inflater we reserve new ringbuffer when table size is
changed.  This may be potentially a problem if new table size is very
large number.  When inflater is not used directly by application, this
is not a problem because application can choose the buffer size.  On
the other hand, if application uses inflater directly and it does not
have control of new buffer size (e.g., protocol dissector), then we
just fail to allocate large buffer in
nghttp2_hd_inflate_change_table_size() without actually use such huge
buffer.  This change defers the actual allocation of buffer when it is
actually needed so that we will fail when it is absolutely needed.

It looks like setting read-rate and read-burst to 0 makes busy loop.
It seems a bug.  On the other hand, we most likely want per-thread
rate limit rather than per-connection.  So we decided to drop them.