Fix GH-23459: Improve imagegrabscreen() and imagegrabwindow() performance on Windows

[LamentXU123]

The original implementation of imagegrabscreen is super weird. It currently calls GetPixel() once for every pixel after capturing the desktop (???). On a 1560x1040 display, this results in more than 1.6 million GDI calls and takes around 2.8 seconds locally. (?????) Besides, I find a out-of-bound off-by-one read in the original implementation. This is also true for imagegrabwindow. They share the same exact logic and don't even make this out as a shared helper...

This is a easy fix. Anyways, now we retrieve the bitmap in a single GetDIBits() call as a top-down 32-bit DIB, then convert the returned RGBQUAD buffer directly into the GD true-color image.

I run the benchmark locally with a 1560×1040 pic.

Before: 2.997434s
After: 0.003682s

A 814x optimization in total. Fixes #23459

[devnexen]

Not out of bounds from what I see ... The y <= Height / x <= Width bound is a genuine off-by-one and should have been <, but neither side of the loop body touches memory out of bounds: if you think I m wrong please add a test.

[LamentXU123]

Blerg I made a typo here. Yeah what I meant is off-by-one :)

[staabm]

should this also mention imagbmp from #23457 ?

[devnexen]

nope. already fixed and different issue anyway (as in its fix can t apply here).

[LamentXU123]

I think it is reasonable to document the optimization in #23457 in the UPGRADING file. But that's a different story. I will take care of this in a different commit.