Hi, I'm student researcher in deep learning and have fair amount of experience in the workings of deep neural networks. I've worked with Vanilla-VAE, Conditional-VAE, and Beta-VAE, in the past. I've also worked on Bayesian Neural Networks as well which also uses variational inference, ELBO loss, kl-divergence, etc.
I went through the notebooks attached. Currently your code doesn't work on cpu as it has grad_scaler (which is meant for gpu), also it doesn't work on gpu either (because the tensors are not at appropriate places). So, it seems that the code has quite some bugs.
I'll be able to fulfil all your requirements. But regarding correspondence of latent space plot of 1D-CNN (pytorch) to 1D-CNN (tensorflow), you may expect different results with Vanilla-VAE and Beta-VAE. Tensorflow version is currently using Beta-VAE. So if you want Beta-VAE in PyTorch, that'll again take some time as it is currently not implemented in PyTorch. Therefore, I might need a couple of days more.