The framework opus decoder discards some samples after a call to
flush(). Because we flush a decoder that is being retained across an
input format change, this means that the start of audio gets truncated
when transitioning to a new opus stream. See also
https://android.googlesource.com/platform/frameworks/av/+/refs/heads/android10-release/media/libstagefright/codecs/opus/dec/SoftOpus.cpp.
Avoid this by recreating opus decoders instead of flushing them. It
seems fine to do this for all opus decoders as reinitialization should
be cheap, OEM-provided implementations may also discard samples and
playback shouldn't be interrupted on reinitialization due to the
downstream AudioTrack buffer.
PiperOrigin-RevId: 277458759
