Fix artifacts regression on AMD Polaris GPUs

[TacoGuyAT]

Workaround #61415 by reverting #98701

Before:

After:

[darksylinc]

This bug is very specific to Polaris cards on Windows drivers.

It seems to be very sensitive to submission order, rather than anything else. I suspect any random thing could break it again.

But if the results of this workaround works consistently enough, I suggest this fix should only be applied when detecting the affected cards / drivers, rather than applying it unconditionally to all systems, and tag it as a workaround so it can ve properly tracked.

(That is as long as validation sync isn't failing)

[TacoGuyAT]

I'm not familiar with Godot's codebase and C++ in general, if there is some existing code I can reference to implement device/driver detection and tailor rendering logic to specific device group - I could implement it myself, but otherwise it would require someone else's help.

[Zireael07]

It might be worth it to have a comment mentioning those drivers/device so that it doesn't get broken again

[DarioSamo]

Further context needs to be provided here that this is specifically rolling back the contents of this PR, which provides tangible performance benefits (#98701). Please don't skip on the context as it makes it more confusing for reviewers to evaluate.

Given we haven't seen it cause issues anywhere else but this family of GPUs on this particular driver, what should be addressed first and foremost is verifying if there's actual synchronization layer validation issues that are detected and fixed by reverting the PR. If that's not the case, then we must only add a workaround only under this specific hardware, as it'd kill the performance benefit of the PR.