Shallow ensemble as an uncertainty quantification method

[Benzoin96485]

1. Shallow ensemble implementation in the basic DenseLayer, installed as a layer hyperparameter in PhysNet and SpookyNet, with atomic property affine, electrostatic energy, charge conservation, dipole monent calculation, energy reduce, force calculation layers supported.
2. A highly flexible shallow ensemble reduce layer is added to collect means and variations from ensemble predictions.
3. Two experimental types of NLL loss based on ensemble means and variations.
4. "In situ" re-evaluation of test set.
5. New monitor output format.
6. Fix the bug that true labels are not inverse-transformed in saved evaluation results.
7. Saving non-target prediction from the raw output of the model is supported