Species-dependent radial basis

[cortner]

With #110 it is now easy to make a Rnl basis, or more generally a scalar basis P_v of the form

P_v = sum_n U_vn J_n

where v is a general multi-index. The parameters U_vn can then be learned as part of the parameter estimation for example.

But we also want Rnla bases where a is the species of the neighbouring atom!! This is a problem because the species is not an index but it is part of the argument.

But is it really a problem? Or is the 1-hot embedding designed to precisely overcome this? So this may or may not really be an issue, but I'm now posting this here to make sure somebody thinks this through and implements an example of such a Rnla basis.