More flexible radial basis embedding

[zhanglw0521]

This PR aims to enable a more general radial embedding, as is addressed in Issue #11

[zhanglw0521]

Looks like this PR is ready to be reviewed. Now the input radial could be an abstract chain, along with its radial specification. The chain itself and its specification are stored in a new structure called Radial_basis (we can at any time change it). The only requirement to the chain is that it should map a configuration Vector{Positions,N_particle} to a $N_{particle} \times N_{invariant}$ matrix, which is rotation invariant.

[zhanglw0521]

Also, I enabled the rSH-based chain (for L=0 only) in this PR (This shouldn't have been in this PR but I saw rSH was used in forces.jl so it might do no harm to just add a few lines in this regard).

[cortner]

Thank you, Liwei. @CheukHinHoJerry - can you do a first review please. I'll then take another look before merging.

[CheukHinHoJerry]

Just to put a note here that 9321a06 resolved reverse over reverse of single species model.

[cortner]

@CheukHinHoJerry -- can you quickly confirm what you had to change?

Is it simply replacing the comprehension with a broadcast?

[CheukHinHoJerry]

Yes - more precisely, I changed from this:

Radial_basis(Chain(trans = WrappedFunction(xx -> [f_trans(norm(x)) * f_cut(norm(x)) for x in xx]),
                poly = lux(basis), ), spec)

to this:

Radial_basis(Chain(getnorm = WrappedFunction(x -> norm.(x)), trans = WrappedFunction(x -> f.(x)), poly = lux(basis), ), spec)

and similarly this also works:

Radial_basis(Chain(trans = WrappedFunction(x -> f.(norm.(x))), poly = lux(basis), ), spec)

Still, I don't understand fully the reason behind this yet.

[CheukHinHoJerry]

I used two work around, which should be taken care by ACEsuit/Polynomials4ML.jl#67 (comment) and #18