Add support in Ridge for a model with different weights per central-species

[agoscinski]

Problem explanation

In the case we cant to build a model with different weights per central-species, we need to build the covariance matrix different. The covariance matrix becomes instead of one n_properties x n_properties block, a block diagonal matrix with n_properties x n_properties blocks. Each block represents a different central species. For example for a dataset with only hydrogen and oxygen species, the covariance matrix would look like this

$$\begin{pmatrix} H\times H & 0\\ 0 & O\times O\end{pmatrix}$$

where the $H\times H$ and $O\times O$ are both n_properties x n_properties blocks.

Proposed solution

In these cases the central species information is stored in the keys of the TensorMap. However for Hamiltonian learning, the different keys are used for learning different entries of the Hamiltonian. The logic is different: the central-species model learns the same property for different keys (energy and its gradients) the Hamiltonian model learns different properties for each key (different entries of the Hamiltonian). Therefore we can automatically differ the two use cases from the target $y$. For the Hamiltonian learning case, the target $y$ should have the same keys as $X$ in the TensorMap. For weights per central-species $y$ still only has one key.

Possible issues with proposed solution

What is still a bit problematic, for now in Ridge we assumed that for total energy predictions the reduction of the feature matrix X to structure features happens before they are passed to the Ridge class. I am not sure if this works in the case of a model with weights per central-species. But here we need to look at the math in more detail.