Separate BNN modules for MLP, ConvNet, RNN, GNN, etc.

[ziatdinovmax]

Currently, BNN, HeteroskedasticBNN, PartialBNN, and HeteroskedasticPartialBNN automatically handle different neural network architectures. While this works well for MLP and ConvNet cases, expanding this approach to other network architectures, such as RNNs and GNNs, may be challenging. Therefore, it may be better to have a separate Bayesian module for each architecture: BayesianMLP, BayesianConvNet, BayesianRNN, etc. Within each module, we can address both homoskedastic and heteroskedastic scenarios.