Self-contained preprocessing code

[brentyi]

Preprocessing for training is a recurring issue: #11, #12, #14.

The cause of this is that our method was originally trained using data processed for another project, where there was some historical complexity. Including:

• A fork of the HuMoR preprocessing script integrated in a much larger system (which had a Hydra-based YAML config and other complexities)
• A custom version of the SMPL-H model (gender-neutral, with hand PCA information concatenated)
• Logic for converting male/female SMPL-H parameters to gender-neutral ones
• Possibly other changes or forks I'm not aware / that have been lost to the sands of time

I haven't been thorough enough with all of the details here. This PR aims to correct that.

To-dos:

• Self-contained version of the HuMoR preprocessing script
  • Updated dependencies
  • Monkeypatch smplx package to prevent shape + PCA errors
  • Tested
  • Gender conversion utilities in repository
• Document in README
• Check whole pipeline end-to-end