Filter sinks that have an internal buffer add the internal buffer size to the rewind request when propagating the request to the master sink. Instead of making the request larger, the request should be made smaller, analogous to how pa_sink_input_request_rewind() handles the data in the render_memblockq.

This probably doesn't cause user-visible problems, because the rewind requests being larger than necessary only causes some extra rendering work and thus higher cpu-use, the audio contents are not affected.