Fully homomorphic encryption, in which you can do arbitrary computation on encrypted data, is still quite slow. But partially homomorphic decryption, in which you can add encrypted values together but not multiply (or vice versa), is quite efficient. And since the secure aggregation protocol only needs to add together encrypted values to get an average, it only needs partially homomorphic encryption properties.
I believe there is also a proof that says any partially homomorphic system can be reworked into a FHE.
You're thinking of "somewhat homomorphic encryption", which is homomorphic encryption that can support both addition/OR and multiplication/AND, but only in circuits of a limited depth. The original FHE paper did indeed prove that you can rework any "somewhat homomorphic" system into a fully homomorphic one.
Partially homomorphic encryption is different because it really only enables one of those two types of operations. For example, Pallier encryption has the property that Enc(A) + Enc(B) = Enc(A+B), but there's no way to go from Enc(A) and Enc(B) to Enc(A×B).
Are they actually using this in this intended federated learning plan? If so that's a truly major innovation.