These could have been nicely separated into per-device translations
(one for GDC, one for EGC, one for regular VRAM accesses)… if their
original order followed that classification.
OK, so let's make it one full `graph_ex.cpp` translation unit… nope,
FUUIN.EXE needs graph_2xscale_byterect_1_to_0_slow() with different
compilation flags. 🤦
Whatever, fuck it, let's go for individual translation units, with
individual headers, so that we at least get predictably organized
source code out of this.
Part of P0196, funded by Yanga.