- load in features if already computed
- allow computing features for multiple fastas in parallel
- allow disabling Amber relaxation, b/c our sequences have X's

sorting fastas by seq len and padding to round up to nearest 50, so
many consecutive sequences can be run by the same compiled Jax model.

Note that different models (1 through 5) have different sizes
(e.g. extra msa size), so for now you must choose exactly one model
in run_docker.py. In the future, may want to refactor run_alphafold.py
so the outer loop is over models instead of inputs.

On a related note, tried running multiple inputs on a single GPU in separate child processes,
but that was actually slower than doing it serially.

- split up feature generation vs model running
- run one model at a time, thru all inputs (for Jax compile speedup),
  and then rank models at the end.
- don't run a model if output already exists

…ker re-build on code changes

(workaround for now, in the interest of time)

…s at the end. Do that now.

…nient file

…beddings .pkl is outputted.

Also, when ranking models, grab plddt from the small embeddings .pkl file, not the huge result .pkl file.

…ings