Overflow in PSNR computation

[incenger]

In the function to compute PSNR score, if the two input image arrays of type np.unint8, using (img1 - img2)**2 can overflow. This causes incorrect PSNR evaluation result, especially when calling the function without convert_to="Y" argument. I think this could be fixed by either explicitly converting the input images into float type or adding the data type requirement into the document.

[innerlee]

@Yshuo-Li pls check if there are corner cases, and fix & add unittest if it is true

[Yshuo-Li]

@incenger , thank you for your reminder, I will fix this issue later.
Note: In this repo, image arrays send to this function are of type np.float32. This issue does not affect the normal use of the repo.

[incenger]

@Yshuo-Li Yes, I only found this issue when using the function on my own code :D

[Yshuo-Li]

@Yshuo-Li pls check if there are corner cases, and fix & add unittest if it is true

Fix this issue in "Tiny fix of PSNR" #294