This assumes that memory for the system and the graphics card is shared.

Yeah, Windows in this era already had the concept of bitmaps in system memory and bitmaps in device memory, so the desktop background could have been decompressed into GPU memory and then thrown away to free up CPU-side RAM. Not sure whether it would actually do that though.

Or, that desktop rendering is not GPU accelerated.

But the uncompressed data doesn't need to stick around. It could be uncompressed piece by piece into a much smaller temp buffer, with the revealed parts of the areas of interest copied into video RAM as necessary.