Stochastic Bregman Proximal Gradient Method Revisited: Kernel Conditioning and Painless Variance Reduction