SSIM returning wrong values and limited to uniform kernel only

[Warvito]

Describe the bug
The SSIMMetric is reshaping the values with view(ssim_value.shape[1], -1) instead view(ssim_value.shape[0], -1) returning a vector of (channels, 1) shape instead (batch, 1).

MONAI/monai/metrics/regression.py

Line 385 in 8eceabf

ssim_per_batch: torch.Tensor = ssim_value.view(ssim_value.shape[1], -1).mean(1, keepdim=True)

Besides that, in most packages we compute the SSIM using an gaussian kernel, instead a uniform kernel. The option to be able to choose the type of kernel should be added.

Finally, in most package, there is no cov_norm parameter, producing a significantly different value when compared to the torchmetrics package with an uniform kernel, and compared to torchmetrics, piq and pytorch-msssim when using an gaussian kernel