RandONets: From Universal Operator Approximation with Randomized Features to Fast and Efficient Least-Squares Training with Decoupled Branch-Trunk Nets