[yantis]

With a v4 (XPS-based) printer driver, Windows converts each bitmap SumatraPDF blits into a PNG in the XPS spool file. A 32bpp band becomes an RGBA PNG even though every pixel is opaque, and the driver renders RGBA images more slowly than RGB ones.

This passes each rendered band through the existing PixmapToBgr() before BlitPixmap(), so the driver gets 24bpp bits and the spool holds RGB PNGs. Print renders carry no transparency (keepAlpha and transparentBackdrop are off for print), so the pixels are the same. 8bpp palette DIBs come back from PixmapToBgr() unchanged.

This changes the GDI path (-print-to, the classic dialog, selections). The Windows 11 dialog renders through Direct2D and is not changed.

Measurements. Master at 80d7528, x64 Release. A 9-page Letter PDF (text and photos) printed with -print-to at 600 dpi to an HP Color LaserJet M555 queue with HP's v4 driver ("HP Color LaserJet M554-M555 PCL-6") on a FILE: port. The queue was paused while SumatraPDF spooled and resumed afterwards, to time the two parts separately (2 runs each):

	SumatraPDF spooling	driver rendering
master	36.7 s, 49.7 s	32.0 s, 21.8 s
this PR	36.3 s, 33.3 s	8.0 s, 11.9 s

Sending each spool file to the queue again unchanged, 5 times each: master 20.3-32.0 s (median 26.3 s), this PR 13.5-25.3 s (median 14.3 s).

Printing normally (the driver renders while SumatraPDF spools), whole job: master 41.1-47.9 s (3 runs), this PR 38.9-41.3 s (2 runs); on this machine spooling hides most of the driver's time.

With -print-settings stretch (bands go through StretchDIBits instead of StretchBlt), whole job, one run each: master 74.3 s, this PR 39.2 s.

Output is unchanged. At 1:1, each spool holds 66 band images of 5100x822: the 65 RGBA bands in master's spool (alpha 255 everywhere) are RGB in this PR's spool with identical pixels, and the one 8-bit palette band is the same in both. With stretch, all 64 band images are identical as well.

Related: #373. With 3.6.1, which sends one full-page 32bpp bitmap per page, this document's spool took 229 s to render on the same driver, and 7 s with only the alpha channel removed from its PNGs (one run each).

[kjk]

Merged, thanks! The spool measurements made this easy to evaluate.