Wouldn't the 3 shears method take a preposterous¹ amount of memory if you rotate by something like 180 degrees? The first shear will take a w.h image and output a w.h.h one!

Why not simply use Bresenham's line drawing algorithm to "draw" a reverse-rotated rectangle over your image, but rather than drawing, you'd read pixels and write them to the destination image. You'll need to only rotate the corners of the image once and everything else is simple integer math.

¹ preposterous for the kind of device where doing it any other way is too slow

Edit: replaced stars with dots for multiplication because stars make italic text