To make the thing really "modular," I'd want the communication channels between cells (both the "I'm alive/dead" channel and the "clock tick" channel) to be something like sliding pegs that push into holes on the neighbors and can be sensed either mechanically or optically, not something that requires an electrical current to flow _between_ neighboring cells.
Adafruit sells a 4x4 Game of Life circuit board, which can be connected (permanently, with solder) to other 4x4 boards to make larger boards: https://learn.adafruit.com/game-of-life Having to actually solder it to the other boards makes it not quite what I'm thinking of, but still pretty darn close.
Re counting 2D totalistic CA rules: it could be as many as 2^8 = 256 rules for "I come alive with (1,2,3,4,5,6,7,8) neighbors," times another 256 for "I stay alive with (...) neighbors," for a total of 63356 possible CAs. One _could_ add 16 switches to each cell, and have the cell use whatever rule is encoded by those 16 switches. That feels expensive, though, doesn't it?