Link or put multi-GPU training docs inside SOAP-BPNN docs

[PicoCentauri]

There is an advanced concept page for multi-gpu training of the SOAP-BPNN, but this is not shown in the architecture docs of the SOAP-BPNN.

At least one should link the page inside the architecture docs but I think better is to put this page directly in the docs of the SOAP-BPNN because this style of multi-gpu training only applies to this architecture.