Add split_by operation to split according to one of the Labels dimension

[Luthaf]

This would allow to do something like split_by(tensor, axis="samples", dimension="structure") to split along structures

[jwa7]

Good idea! The current API equivalent, though certainly more fiddly, is:

import numpy as np
import metatensor as mts

tensor = mts.load("path/to/tensor.npz")
split_by_axis = "samples"
split_by_dimension = "structure"
grouped_labels = [
    mts.Labels(names=split_by_dimension, values=np.array([A]))
    for A in mts.unique_metadata(tensor, axis=split_by_axis, names=split_by_dimension)
]

split_tensor = mts.split(tensor, split_by_axis, grouped_labels)

I guess the easiest way to implement split_by is to wrap the private functions in split.py and using unique_metadata