Use these instead of exposing the internal roar_t struct.

Copy the reason_phrase from the SoupMessage, and check for transport
errors in the session callback.

Add GMutex, GCond attributes which will be used by callers to
conditionally wait on the stream.

Remove the (now-useless) plugin method buffer(), wait on GCond
instead.  Lock the input_stream before each method call.  Do the same
with the playlist plugins.

To check for errors without reading.  The decoder thread wants to do
that, before it passes the input stream to the plugin.

Fix memory leak.

libsoup's asynchronous API is not thread safe.  By moving the calls
into the I/O thread, several crash bugs will be fixed.

Add the "complete" attribute, and set it in _session_callback().
_close() waits for it to become true.

Add attribute postponed_error, pass this GError to the caller.

Fix deadlock.

We'll add proper locking now, and there's no need to carry this
duplicate flag.

To report failures to callers who don't use _buffer().

Avoid recursive deadlock in _session_callback().

This was useless, because input_stream_read() blocks until data is
available.  Calling it in a loop doesn't make sense.

Pausing and resuming after every little chunk adds lots of overhead.
Add a lower level for resuming the stream.

curl_easy_pause(CURLPAUSE_CONT) may have added enough data to go into
pause again.

Call _update_fds() directly.  This is possible because it's only
called from within the I/O thread.

Fix deadlock.

Add -DNDEBUG to AM_CPPFLAGS.

Conflicts:
	NEWS
	configure.ac

Let the codec start with fresh buffers.  This should fix the remaining
seeking issues.

Use AVStream.time_base to convert the decoder_seek_where() value, and
pass the current stream number to av_seek_frame().

Use flag AV_TIME_BASE.

Code simplification.

Reduce heap usage by reducing the number of malloc() / free() calls.

The global data structures are now lock-free, because they are
accessed only from the I/O thread.  By using per-request locks, we
have finer grained locking, preparing for locks shared with the
client.

This adds some overheads for indirect calls to the I/O thread, but
reduces the amount of global locks.  Next step will be switching to
per-request locks.