Add inference_only option to trainers

[lbluque]

Add inference_only option to trainers, so that datasets, optimizer, and other objects not necessary for inference are not loaded from checkpoints.

This is mostly for the OCPCalculator, and makes resolves a handful of issues that otherwise arise:

• no need to delete datasets anymore since these will not be loaded
• creating the trainer will not look for files that were used for training but are not necessary for inference (such as normalizers which are already part of the checkpoint)
• no need to specify the specific trainer when loading a checkpoint in OCPCalculator, the default OCPTrainer should work almost all of the time.

And deprecate OCPCollater in favor of using a partial to set otf_graph.

[misko]

Looks good! LGTM!

[codecov]

Codecov Report

Attention: Patch coverage is 81.25000% with 9 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/fairchem/core/trainers/base_trainer.py	79.06%	9 Missing ⚠️

Files with missing lines	Coverage Δ
src/fairchem/core/common/data_parallel.py	77.50% <100.00%> (-1.50%)	⬇️
src/fairchem/core/common/relaxation/ase_utils.py	61.95% <ø> (-0.81%)	⬇️
...e/models/equiformer_v2/equiformer_v2_deprecated.py	91.15% <ø> (ø)
src/fairchem/core/trainers/ocp_trainer.py	69.49% <100.00%> (ø)
src/fairchem/core/trainers/base_trainer.py	86.12% <79.06%> (-2.20%)	⬇️

... and 2 files with indirect coverage changes