[BUG] SSIM returns nan for sidelength < 11

[stroblme]

Describe the bug
Calculating the SSIM metric for target/pred with sidelength <11 returns nan.
The reason is, that the default Gaussian Kernel uses +5 pixel in each direction thus requiring an image of at least 11 pixels.
The kernel size is automatically calculated (kernel_size parameter only considered for a uniform kernel) based on the sigma value.
E.g. sigma=1.0 produces a gaussian kernel of 9x9, thus allowing for images with sidelength >=9.
Reducing sigma would help, but raises the question if such a kernel still produces a meaningful SSIM score.

To Reproduce
Steps to reproduce the behavior:
Run experiment with sidelength <11, observe the reported SSIM metric.

Expected behavior
Either raising and error or adjusting the kernel size.

Screenshots
If applicable, add screenshots to help explain your problem.

Hardware
Please include some information regarding the hardware you've used.

Additional context
Add any other context about the problem here.