Currently, resume_read() fails if on_read() returns -1 in case that
evbuffer_add failed, which means, most likely, memory allocation
failure. ClientHandler is marked "should be closed", but if
evbuffer_add is failed, write callback will not be invoked and its
marking is not evaluated. It will eventually be deleted when the
client is disconnected or backend failure though.

Specify proxy URI in the form http://[USER:PASS]PROXY:PORT. USER and
PASS are optional and if they exist they must be properly
percent-encoded. This proxy is used when the backend connection is
SPDY. First, make a CONNECT request to the proxy and it connects to
the backend on behalf of shrpx. This forms tunnel. After that, shrpx
performs SSL/TLS handshake with the downstream through the tunnel. The
timeouts when connecting and making CONNECT request can be specified
by --backend-read-timeout and --backend-write-timeout options.

With --spdy-bridge option, it listens SPDY/HTTPS connections from
front end and forwards them to the backend in SPDY. The usage will be
written later. This change fixes the crash when more than 2
outstanding SpdyDownstreamConnection objects are added to SpdySession
and establishing connection to SPDY backend is failed.

This option specifies additional certificate and private key
file. Shrpx will choose certificates based on the hostname indicated
by client using TLS SNI extension. This option can be used multiple
times.

INFO log and its surrounding code are now guarded by
LOG_ENABLED(SEVERITY) macro so that they don't run if log level
threshold is higher. This increases performance because log formatting
is somewhat expensive.

create_error_html() is rewritten without std::stringstream.