Fixes #2913

Description:

This pull request updates the code in ignite/metrics/ssim.py so it won't create extra torch.Tensors.
The changes involve modifying the list comprehension that creates the output_list, which has the wrong length of BatchSize*5 instead of it being 5 (the same as len([y_pred, y, y_pred * y_pred, y * y, y_pred * y])). The elements with index>4 are empty tensors with shape (0,C,H1,W1) in the output_list

Instead of concatenating the tensor inputs in torch.cat([y_pred, y, y_pred * y_pred, y * y, y_pred * y]), I have created a list of the input tensors as input_list= [y_pred, y, y_pred * y_pred, y * y, y_pred * y], which I believe was the actual intention of the author.
Now we have access to the Length of this list, which we will use to retrieve input_list elements after passing them through a conv2d layer.
I then passed the concatenated input tensor using torch.cat(input_list) which has a size of (B*5, C, H1, W1) to the F.conv2d function call.

After that, the outputs which is a tensor with shape (B*5, C, H, W) is sliced for every Batch element to return a list of 5 elements. this is where the previous code had a problem and instead of a list of 5 (or len(input_list)) tensors, it was returning a list of B*5 elements, where only its first five elements were actual tensors and the rest where empty when the Batch_size>1.
(the previous code was calling tesnor[index1:index2] where index1 and index2 were larger than tensor.size(0) which was returning empty tensors)

We can see in lines 167-173 that the elements with indices 0-4 are being used, which proves this issue.

I have tested these changes on my local machine and confirmed that they do not introduce any new issues. Please let me know if there are any concerns or feedback regarding these changes.

Thanks for considering this pull request!

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)