Now config global is backed with std::unique_ptr.  configuration
swapping dance is now a bit cleaner, but YMMV.

This commit also fixes the bug that old configuration is still used
for worker process.  The another bug fix is that inherited, but not
used fd is not closed in worker process.  That makes reloading next
configuration fail if it contains the address which are leaked into
worker process.

The last_worker_pid is known by inspecting the last entry of
worker_processes.

This commit implements configuration reloading with SIGHUP.
There are rough edges left:

* Rename SignalServer with more meaningful name, say, WorkerProcess.
* We should introduce global configuration object which is not
  affected by configuration reloading.  It should hold cmdcfgs, argc,
  argv, and last worker PID.
* We should close the listener file descriptor when some operation was
  failed after that.

The N-th -p options sets N-th URI's weight.  If the number of -p is
less than the number of URIs, the last -p value is repeated.  If no -p
is set, use default weight, 16.

clang + libc++ does not work, it still requires libc++_shared.so
runtime even if -lstdc++ is used, which supposed to link with static
version of libc++.

It was markdown, we should use reST.

Previously, we didn't retry request on connection failure.  Sometimes
we hit the edge case where connection is about to lost just when we
write request.  To avoid this situation, we now retry request to
failed attempt.  We also add ConnectBlocker to MemcachedConnection not
to attempt to connect to memcached if connection could not be made
previously.

We use technique described in https://github.com/RJ/ketama

We define the behaviour when NGHTTP2_ERR_CANCEL is returned from
before_frame_send_callback.  That is to cancel the frame passed to the
callback.

In our use case, x and y is quite small, and there is no chance for
overflow, and y is always integer.