Add train-only runner and latest checkpoint tracking

[bkmi]

Summary

Adds a train-only runner primitive and tightens checkpoint handling after the callback refactor.

Changes:

• Introduce TrainRunner, a core runner backed by TorchTNT train().
• Refactor shared train runner mechanics into _BaseTrainRunner so TrainRunner and TrainEvalRunner share:
  • callback hook wiring,
  • checkpoint save/load,
  • preemption resume handling,
  • timeout callback setup.
• Keep TrainEvalRunner as the TorchTNT fit() runner and preserve its existing train_eval_unit interface for compatibility.
• Update TrainCheckpointCallback to:
  • maintain a latest symlink for step and final checkpoints,
  • rotate saved checkpoint directories without deleting latest,
  • validate max_saved_checkpoints >= 1,
  • remove unused stored load_callback state.
• Add W&B tag passthrough to WandBSingletonLogger.init_wandb.
• Add focused tests for train-only runner behavior and checkpoint rotation/latest handling.

Why

Some workflows need train-only execution without eval dataloaders or validation/test loss computation. This adds that capability as a fairchem core primitive instead of keeping it model-local, while preserving the existing train/eval runner behavior.

The latest-checkpoint symlink also gives downstream jobs and operators a stable checkpoint path while still retaining bounded checkpoint history.

Testing

Ran locally:

pytest tests/core/units/mlip_unit/test_train_eval_runner.py -q --rootdir .
pre-commit run --files \
  src/fairchem/core/components/train/train_runner.py \
  src/fairchem/core/components/train/callbacks.py \
  tests/core/units/mlip_unit/test_train_eval_runner.py
python -m ruff check \
  src/fairchem/core/components/train/train_runner.py \
  src/fairchem/core/components/train/callbacks.py \
  tests/core/units/mlip_unit/test_train_eval_runner.py \
  --select F401,F841,F821,ARG