Doomformer: compiling an integer neural network into a DOOM map, using voodoo dolls to do arithmetic

Doomformer compiles an integer neural network into a ZDoom map and runs it in ViZDoom, with voodoo dolls performing arithmetic and sector lighting and floor heights storing values; Python reads back each value from the running game and compares it against a NumPy integer reference with a tolerance of 0. It is about 23,000 times slower than NumPy; a handwritten digit takes 163 seconds in real-time DOOM under the default schedule and 8 seconds with --opt 3. On 10,000 MNIST test digits, the integer network achieves 97.16% accuracy, and all results read back from DOOM are exact.