If the output is already open, the `current_chunk` pointer may be
bogus and out of sync with `SharedPipeConsumer::chunk`, leading to an
assertion failure in `SharedPipeConsumer::Consume()`.

Fixes #411

The output plugin shall decide whether to insert silence or do nothing
at all.  The ALSA output plugin has already implemented this.
Inserting silence is not necessary or helpful for some plugins, and
may even hurt them (e.g. "recorder").

This will be used by filters which have internal buffers which need to
be flushed at the end, e.g. the "soxr" resampler.

This fixes spurious replay gain logs when the player inserts silence
chunks, because those silence chunks had no replay gain attached,
resetting the ReplayGainFilter state, flipping it forth and back.

This eliminates some overhead, because the compiler doesn't need to
consider these functions throwing.

Each close/open cycle resets the Filter's state, because a new Filter
instance is being created.  That results in the serials
(replay_gain_serial and other_replay_gain_serial) being out of sync
with the internal ReplayGainFilter state.

So instead of initializing those serials once, we need to initialize
them each time we create new ReplayGainFilter instances, i.e. in
OpenFilter().

 https://bugs.musicpd.org/view.php?id=4632

Fixes assert failure after unpausing.

The filter can take some time to finish, even more so on a weak
machine with a saturated CPU.  By not holding the mutex during that
time, we can reduce PlayerThread latency a lot, because that thread
needs to synchronize a lot with all outputs.

Don't expose MusicChunk instances, provide higher-level access to
chunk contents.