This fixes a problem which caused a failure with snd_pcm_writei()
because snd_pcm_drain() had already been called in the previous
iteration.  This commit makes sure that snd_pcm_drain() is only called
after the final snd_pcm_writei() call.

This fixes discarded samples at the end of playback.

When a playback error has occurred, MPD would never recover until one
restarts MPD.

MPD's default is 100ms, which is too long for the real-time I/O
thread.  The OutputThread has 100us, but the real-time I/O thread
might have tighter deadlines.

This change has currently no effect (I believe), because nobody uses
timers on the RTIO thread.

CancelInternal() doesn't need to be protected because it is called
synchronously from Cancel().

Fixes a theoretical race condition which could occur in Drain() (but
was extremely unlikely).

If our `ring_buffer` is smaller than the ALSA-PCM buffer (if the
latter has more than the 4 periods we allocate), it can happen that
the start threshold is crossed and ALSA switches to
`SND_PCM_STATE_RUNNING`, but the `ring_buffer` is empty.  In this
case, MPDD will generate silence, even though the ALSA-PCM buffer has
enough data.  This causes stuttering (#420).

This commit amends an older workaround for a similar problem (commit
e08598e7) by adding a snd_pcm_avail()
check, and only generate silence if there is less than one period of
data in the ALSA-PCM buffer.

Fixes #420

The method Cancel() assumes that the `period_buffer` must be empty
when `active==false`, but that is not the case when Play() fails.

Of course the assertion in Cancel() is not 100% correct, but I decided
to rather fix this in LockCaughtError() because the `period_buffer`
should only be accessed from within the RTIO thread, and this is the
only code path where `active` can be set to `false` with a non-empty
`period_buffer`.

Fixes #423

This implements real error handling, and avoids calling
CancelInternal() from this code path.

alsa-lib doesn't set errno, it returns errors as negative integers.

This check was added 9 years ago in commit
4dc25d39 to work around a dmix bug
which I assume has been fixed long ago.

Removing this fixes another corner case: if draining is requested
before the start threshold is reached, the PCM is still in
SND_PCM_STATE_PREPARED but not yet SND_PCM_STATE_RUNNING, which means
the submitted data will never be played.  This corner case is
realistic when playing songs shorter than the ALSA buffer (if the
buffer is very large).

This fixes a corner case which has probably never occurred and
probably never will: if Cancel() is called, and then Play() followed
by Drain(), the plugin should really play that data.  However
currently, this never happens, because snd_pcm_prepare() is never
called.

Closes #410

I added this sentence in commit
5271e81e, but this was merely
documented the legacy status quo, which has always been undocumented
for old-style filters.

But for new filters, using "==" for sub strings was a surprising
"feature", which I removed in commit
ac0852b4.

This fixes the assertion failure due to calling
AudioOutputSource::Close() twice.

Fixes a crash bug with `always_on` outputs which occurs because the
`AudioOutputSource` still has a pointer to an outdated `MusicChunk`.

Fixes #415

With the new command, the decision to pause or close the output moves
into the output thread.

This feature requires `libpcre`.

This call was missing, causing very high CPU usage when the ALSA
output plugin was used with dmix.

Closes #391

When `metadata_sent` is `false`, the plugin assumes there is metadata
which must be sent, even if no metadata page was passed to the plugin.
Initializing it to `true` avoids dereferencing this `nullptr`.

Fixes #412