debug for the inference of gray image

[DsLipku]

#483
repeat the gray channels to three channels before use GFPGAN to enhance face.

[Splendide-Imaginarius]

I can confirm that 08bb631 fixes the error. However, I suspect that this fix should be submitted to GFPGAN rather than this repo, so that other downstream projects benefit from the fix. Also, I do not understand why img_mode is set to 'L'. That variable doesn't look like it's used anywhere except checking for 'RGBA', and I'm pretty sure its value will never be 'RGBA' when the grayscale code path is reached, so I think that line can simply be removed from this patch.