fix: pixman stride bug with bgr888 formats

[cashmeredev]

fixes #47

the screencopy protocol can return buffers with non-aligned strides, especially with bgr888 (24-bit) formats. this breaks pixman since it expects stride in uint32_t units and aligned to 4 bytes.

for non-32bit formats or misaligned strides, we now convert directly to argb8888 by writing into pixman's buffer instead of passing the raw buffer.

tested on nixos + swayfx + nvidia rtx 3070 ti where it was failing consistently before.

[moverest]

Thank you for your changes and your interest in the project. I'm not really sure how to reproduce this on my machine. Does this happen with grim as well? I don't remember them having specific code for this format so I am curious to know if it also fails there.

[cashmeredev]

the crash is caused by how wl-kbptr passes the screencopy buffer to pixman. the compositor can return 24‑bit bgr888 buffers where the row stride in bytes is not a multiple of 4. pixman expects any external buffer to have a 4‑byte‑aligned row stride (it stores the stride internally in uint32_t units), and if that is not true, it hits the check
rowstride_bytes % sizeof(uint32_t) == 0 and aborts.

previously, wl-kbptr just handed the screencopy buffer directly to pixman, so on setups where the compositor used bgr888 with a “weird” stride, pixman would immediately crash. with grim i don’t see this issue, presumably because it only ever gets formats/strides that already match pixman’s expectations.

on my side i can reliably reproduce and compare the behavior:

• nvidia workstation, swayfx, started with --unsupported-gpu
  • old wl-kbptr build: crashes with the pixman assertion about rowstride_bytes % sizeof(uint32_t) == 0.
  • patched wl-kbptr (this pr): works fine, no pixman abort.
• intel laptop (no --unsupported-gpu, integrated gpu)
  • wl-kbptr always works here; i cannot trigger the crash with either the old or the patched build.

this matches the idea that certain compositor/gpu combinations (in my case nvidia + swayfx + --unsupported-gpu) return 24‑bit bgr888 screencopy buffers with a non‑aligned stride, which pixman then rejects. this pr fixes those cases by converting such buffers into an internal argb8888 image with a proper 4‑byte‑aligned stride before passing them to pixman.