Refactors the video pipeline generation from video.py to pipeline_builder.py

[xsdg]

This is not a 1-to-1 refactor. There was a lot of not-clearly-intentional
inconsistency in the original implementation which was not all carried over.
Among other changes:

• The original implementation used the veryfast preset in some places, and
  ultrafast in others. Those presets are described at
  https://trac.ffmpeg.org/wiki/Encode/H.264#a2.Chooseapresetandtune
  The new implementation uses veryfast across the board.
• The logic around resolution, user_command, input_file, and subtitles was
  extremely confusing. This has been significantly normalized. I wouldn't be
  surprised if this results in regressions, but we'll be better off fixing those
  regressions in the new implementation rather than trying to figure out how/why
  they worked in the old one.
• Argument placement in the generated commands will be much more stable across
  input conditions than they were in the legacy implementation. This may cause
  regressions in cases where ffmpeg was sensitive to the specific order of those
  arguments in different conditions.

This refactor makes significant progress on all three subgoals of #434. Most
importantly, there is no longer any work done at import time when importing
video.py.