fix: improve vae stability by controlling the minimum std that can be learned

[PabloNA97]

For some datasets I obtained the following error:

VAE training failed. Error message: Expected parameter scale (Tensor of shape (128, 2)) of distribution Normal(loc: torch.Size([128, 2]), scale: torch.Size([128, 2])) to satisfy the constraint GreaterThan(lower_bound=0.0), but found invalid values:
tensor([[3.5159e-08, 2.5230e-08],
        [1.7620e-23, 3.2989e-25],
        [6.6246e-12, 2.3463e-10],
        [9.5324e-04, 3.7065e-03],
...

This is related to the scale parameter of torch.distributions.Normal being too small. To improve numerical stability during training I think it's a common practice to limit the minimum std that can be learned through torch.clamp().

A minimum of -15 in the log variance would be a minimum value of 1e-15 in the std.

With this modification the instability is gone :)

[PabloNA97]

I forgot to remove the eps (1e+8) that I was adding to the std. This was the first approach but I think torch.clamp should work better.