Change VectorExpansion that metadata is similar to rascaline

[agoscinski]

VectorExpansion has the form

TensorMap with 13 blocks
keys: l
      0
      1
        ...
      11
      12
TensorBlock
    samples (39656): ['structure', 'center', 'neighbor', 'species_center', 'species_neighbor', 'cell_x', 'cell_y', 'cell_z']
    components (1): ['m']
    properties (20): ['alpha_j', 'n']
    gradients: None

rascaline SphericalExpansionByPair has the form

TensorMap with 80 blocks
keys: spherical_harmonics_l  species_atom_1  species_atom_2
                0                  1               1
                1                  1               1
                              ...
                3                  8               8
                4                  8               8
TensorBlock
    samples (9522): ['structure', 'pair_id', 'first_atom', 'second_atom']
    components (1): ['spherical_harmonics_m']
    properties (6): ['n']
    gradients: None

where the self pairs are stored as pair_id -1. (These are two different datasets any hypers that is why number of samples and so on does not match)

• The pair_id is a bit very specific information based on how rascaline systems work. This one I would not try to put into torch_spex
• The cell shift information 'cell_x', 'cell_y', 'cell_z' is not really important anymore since we do the computation of the direction vectors outside of SphericalExpansion. I would just keep it for the moment, but drop it if it becomes necessary
• Having the sparsification on the species centers and neighbors, seems to me smart to do for the nonalchemical part, but for the alchemical part it seems to make sense to keep it in the properties