When using secure aggregation, auxiliary variables are encrypted (thanks to #1007 ) along model weights, but the encryption is done separately.
While it should be ok for LOM secagg algorithm, it is more concerning for JLS, where parameters are encrypted using the same key, and could be deciphered through replay attacks.

The goal of this PR is to encrypt both model weights and auxiliary variable at once, and therefore providing a better security when manipulating model and optimizers parameters.