CalibrationMetrics does not work for datasets with varying numbers of atoms

[lkkmpn]

When lists of Atoms objects with varying numbers of atoms are passed to CalibrationMetrics, the analysis fails. In get_data, converting the lists of forces into arrays will raise an error about the inhomogeneous shape:

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (197,) + inhomogeneous part.

(197 is the number of Atoms objects in the dataset.)

PredictionMetrics works around this by using np.concatenate() instead of np.array(), which is fine because afterwards the arrays get flattened anyway:

true_forces = [x.get_forces() for x in self.x]
true_forces = np.concatenate(true_forces, axis=0) * 1000
self.content["forces_true"] = np.reshape(true_forces, (-1,))

This can be implemented into CalibrationMetrics without an issue, but in preparing the ensemble force data reshaping is done based on the number of atoms in the dataset, and such the above change does not apply there.