fix(canvas): use a corrected gradient mask for canvas pasteback

[dunkeroni]

What type of PR is this? (check all applicable)

• Refactor
• Feature
• Bug Fix
• Optimization
• Documentation Update
• Community Node Submission

Have you discussed this change with the InvokeAI team?

• Yes
• No, because:

Have you updated all relevant documentation?

• Yes
• No

Description

• Gradient Mask Node skips processing and returns tensorified version of image if edge_radius=0
• Gradient Mask Node now outputs an image representing the expanded mask area of the gradient.
• Canvas workflows use this expanded mask for the paste-back operation (goes through MASK_RESIZE_DOWN)

This enables two new use cases in workflows:

• Inpainting graphs can use the output for paste operations in a larger image without losing gradient denoise details
• Pre-made gradient mask images (by hand, depthmap, etc.) can be converted to mask tensors without modification and/or return an image of the total affected region.

This fixes a bug noticeable when using a large edge radius where the canvas paste back would clip the result to the original mask boundaries and then apply a paste operation along that edge. It will now respect the total area affected by denoising and compositing mask blur will be applied to that outer boundary.

QA Instructions, Screenshots, Recordings

Example of the bug: Edge Size 128, Minimum Denoise 0.3, Staged // Mask Blur 0

Merge Plan

This PR can be merged when approved

Added/updated tests?

• Yes
• No

Tagging @blessedcoolant in case of any ideas for a better implementation

[blessedcoolant]

Seems fine to me.