Fixed a bug related to Zero Standard Deviation (Division by Zero error)

[Arka161]

In some cases, if an image consists of homogenous patches, we might get a division by zero issue that results in no features being detected.

To handle this issue, I have added an epsilon, that averts the error. The same fix has been raised in the PyTorch repository of QATM.

A zero SD in the subset of tensors is pretty common in various kinds of industrial images.