Fix regression by commit 45f6129a

See http://bugs.musicpd.org/view.php?id=4477

Instead of stopping playback completely, only CANCEL the queued song
if necessary, and use the SEEK command to play the selected song.
SEEK will take care for current playback state.

Assigning nullptr to next_song may disrupt the player thread and
render undefined behavior.

Another big chunk of code for multi-player support.

From now on, struct Song will be used by the database only, and
DetachedSong will be used by everybody else.  DetachedSong is easier
to use, but Song has lower overhead.

Finally restores an important feature that has been broken for several
months when the PlayerThread started working with Song copies instead
of pointers to the Queue's Song instances (commit e96779de).

Replaces GLib's GError.

This fixes a deadlock bug introduced by 18076ac9.  After all, the
second Cond was necessary.

The problem: two threads can wait for a signal at the same time.  The
player thread waits for the output thread to finish playback.  The
main thread waits for the player thread to complete a command.  The
output thread finishes playback, and sends a signal, which
unfortunately does not wake up the player thread, but the main
thread.  The main thread sees that the command is still not finished,
and waits again.  The signal is lost forever, and MPD is deadlocked.

The main_cond variable was completely unnecessary.  The pc.cond object
can be used for both main->pc and pc->main.

Don't let the "client" pass its own GCond.  This was not used
consistently.

Fixes build problems with mingw32.