The buffer from decoder_NewAudioBuffer() was only used if the audio
output format was planar. Now it is allocated only in that case.

This also remove leaks when decoder_NewAudioBuffer() failed.