args.scale for more than 4x in one pass? about real-esrgan

[xinntao]

@danieljue Thanks for the feedback:-)

1. I think we can train a model for larger scales, such as 8x and 16x. But it may not be very often to use. Could you share your scenarios to use these models?
2. For other upsampling scale factors, like 3x, and 5x, we may add post-processing to quickly achieve it.
3. Currently, -s /--scale option does not work. We will improve it later~

[woctezuma]

I suspect that this comment was removed, but without fixing the issue.

    # FIXME: currenly RRDBNet in BasicSR does not support scale argument. Will update later

Comment added in this commit:
7fcc11f#diff-dbb86c7f0f27adcbccd7badc1a2f2ea8d715383d38737bcf387e3f4f6e87fd12

Then deleted in this commit:
9baeba5#diff-dbb86c7f0f27adcbccd7badc1a2f2ea8d715383d38737bcf387e3f4f6e87fd12

---

It should work with x1, x2 and x4, according to XPixelGroup/BasicSR@406fc6a