LPIPS with Alex net returns Nan

[ilpoli]

LPIPS with Alex net returns Nan

The issue is caused by some of the feature layers might zero and it leads to Nan during their normalization. Original implementation uses small epsilon to avoid the issue https://github.com/richzhang/PerceptualSimilarity/blob/31bc1271ae6f13b7e281b9959ac24a5e8f2ed522/lpips/__init__.py#L13

I've verified that results, after changes, are the same as in the original lib (PerceptualSimilarity).

Here are pictures that lead to Nan measure (I hope GitHub attaches images without additional compressing if so, I'll send them in zip):

[ilpoli]

@francois-rozet could you look at the changes?

[francois-rozet]

Hello @ilpoli, thank you for your PR! I made epsilon an argument of LPIPS. I will merge shortly.