Rework the ffmpeg filter generation in the audio rendering.

This gives the following advantages over the previous code:

* The ffmpeg input filters are loaded from a filter "script" file instead
  of passed on the command line. This fixes some cases of recordings
  failing to process because the ffmpeg command line generated for the
  audio processing exceeded the max command line length limit. (Although
  that only really happens due to BBB bugs...)
* Use absolute positions when trimming audio segments for cuts.
  Previously segments were trimmed to the length of the segment, and the
  results were concatenated. There's some possibility of accumulated
  errors in the segment lengths causing audio desync over time. The new
  code incrementally concatenates the segments, and cuts each segment
  end based on the absolute time since the start of the meeting, to
  avoid error accumulation.