[Feature]: txt2img outpainting feathering/mask blur

[catboxanon]

See #1597 (comment) for background.

txt2img outpainting currently still produces a visible seam if you look closely enough. Depending on the image this can be more or less obvious in some cases. The solution here is also not ideal because it's still a hard cut off, it still fails in some cases, and removes more of the original image that would be kept if feathering were used. The simplest method I see to work around this is employ a feathering technique, like the one already implemented in the webui's img2img inpaint functionality.

The idea would be to allow the inpainted area to be slightly larger by a given amount, as to replace part of the original image, but before the result is returned to the user, the original image would be composited (or "overlaid") on top and have the edges feathered by that same given amount. This is essentially the same thing the webui does for it's inpaint implementation from my understanding.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/f865d3e11647dfd6c7b2cdf90dde24680e58acd8/modules/processing.py#L1196-L1206
https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/f865d3e11647dfd6c7b2cdf90dde24680e58acd8/modules/img2img.py#L117-L125

I plan to try implementing and opening a PR for this myself, but if this issue remains dormant for a few days you can safely assume I did not succeed in my attempt. 😅

[catboxanon]

While working on this I actually now notice the current implementation is already supposed to be doing something similar, just linking here for posterity.

sd-webui-controlnet/scripts/controlnet.py

Lines 904 to 906 in 098f6cd

	sigma = 7
	final_inpaint_mask = cv2.dilate(final_inpaint_mask, np.ones((sigma, sigma), dtype=np.uint8))
	final_inpaint_mask = cv2.blur(final_inpaint_mask, (sigma, sigma))[None]