Make CowData self-inserts safe

[RandomShaper]

When inserting an element taking the original from the current data block so that it will be resized, potentially invalidating the reference to the original, a backup is made to base the copy on it instead.

Fixes #31736.

Other related items (fixed, superseded or plainly related): #16961, #29408, #31159, #31609, #31694.

P. S.: Haven't had time to test this thoroughly.

[akien-mga]

Ah I hadn't noticed that your PR'ed it on the vulkan branch, but it should track master. The vulkan branch is strictly for Vulkan-related changes, I rebase it every once in a while on master.

[RandomShaper]

OK. I'll fix this.

What about changes for 4.0, like C++11 atomics? Is master for 4.0 now?

[akien-mga]

What about changes for 4.0, like C++11 atomics? Is master for 4.0 now?

Not yet, but soon. For now it's probably best to PR such changes against master for review, and redo them once the vulkan branch is merged.

[akien-mga]

For 3.2 we merged #34618 as a simpler stopgap solution, so we'd probably consider that one too for cherry-picking in 3.1.x.

This PR should be reviewed for 4.0. There was a lot of discussion on IRC with @reduz and @hpvb regarding what should be done exactly to solve those issues, which might be worth rediscussing with you.

[reduz]

IMO I would just close this, the solution is bigger than the problem. As I mentioned in the existing fix, the drawbacks of just copying the value are next to none in real life scenarios.