Updating floating point precision for OC22/total_energy predictions and enabling OC22 challenge submission file generation

[wood-b]

This PR does the following:

1. Predictions (energy and forces) for total energy models are written in float32. Ultimately, this only changes the forces_trainer, the energy_trainer was already writing predictions in float32
2. The method used to convert the energy predictions tensor was updated to keep the dtype intact
3. The make_submission_file.py script was updated to include OC22 challenge submissions
4. An assert was added to the make_submission_file.py script ensuring OC22 submission are written in >= float32, this was added because if the total_energy=True flag is not set in the dataset config for an OC22 S2EF model the model will train but predictions will be written in float16 the OC20 default
5. All OC22 base.yml configs were updated to include total_energy=True

I tested OC22 predictions written in both float32 and float16, and the updates work as expected.

[mshuaibii]

Tested make_submission_file.py for OC22. We should go back and try this for OC20 as well.

[codecov-commenter]

Codecov Report

Merging #421 (2d21fd9) into main (cba9fb6) will decrease coverage by 0.01%.
The diff coverage is 14.28%.

@@            Coverage Diff             @@
##             main     #421      +/-   ##
==========================================
- Coverage   45.22%   45.21%   -0.02%     
==========================================
  Files          88       88              
  Lines        8278     8281       +3     
==========================================
  Hits         3744     3744              
- Misses       4534     4537       +3     

Impacted Files	Coverage Δ
ocpmodels/trainers/energy_trainer.py	15.65% <0.00%> (ø)
ocpmodels/trainers/forces_trainer.py	9.57% <16.66%> (-0.10%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[wood-b]

A few updates:

• I ran a number of test predictions with this branch to compare with previous predictions, one model from each of the following tasks: OC20 S2EF, OC20 IS2RE, OC22 S2EF, OC22 IS2RE — sharing those results offline as it is a bit much to add here.
• Documentation for the prediction_dtype has been added to the S2EF example.yml
• Documentation on making an OC22 EvalAI submission file has been added to TRAIN.MD

@abhshkdz @mshuaibii Take another look and let me know if I missed anything.

[mshuaibii]

LGTM - thanks!