Improved parameter detection for Compact

[RunDevelopment]

Changes:

• Removed the pixelshuffle_shape parameter. We are now using the exact model code from the Real-ESRGAN repo.
• Improved detection of output channels and scale. We now support grayscale to RGB compact models.
• Added tests to verify the new detection code.

The new detection code is actually pretty simple: we just brute force it. We know that the number x we read from the state dict is x = scale * scale * out_nc. We also know that out_nc is either 1, 3, or 4. So we just have to try out all possible values for out_nc see which one fulfills the formula for x.

Unfortunately, this method does not work correctly when the actual number of output channels is 1 or 4, because 1 and 4 are square numbers. E.g. for scale=3 and out__nc=4 we get x=36, but scale=6 and out_nc=1 also produces x=36. To fix this, we make the assumption that out_nc=in_nc is likely. E.g. if in_nc=4, we will prefer out_nc=4 over out_nc=1.

[joeyballentine]

Great work