examples: Conway's Game of Life on the ANE

[sbryngelson]

A tiny, very shareable demo in the spirit of examples/reaction_diffusion.py and examples/train_neural_ca.py: Conway's Game of Life with every generation as one on-engine forward pass.

Approach:

• The neighbor count is a fixed 3x3 convolution with a center-zero kernel (all ones except the middle). aneforge/graph.py has af.conv / conv2d (see line ~455) -- use it with a constant weight, pad=1 (wrap-around is optional; zero-pad is fine for a first version).
• The birth/survival rule is elementwise: a cell is alive next step iff (neighbors == 3) or (alive and neighbors == 2). Build it from select / maximum / clip (all in graph.py) or from arithmetic on the boolean-ish masks -- no host loop over cells.
• Roll out N steps, dispatching the step graph each generation, and render frames to an animated webp with PIL exactly like reaction_diffusion.py does (from PIL import Image). Seed with a glider gun or a random field.

Deliverable: examples/game_of_life.py (self-contained, one file, module docstring saying what it shows), plus a one-line entry in examples/README.md. Keep the output asset small.

Testing: compare the first few generations against a trivial numpy Life reference to confirm the on-engine rule matches, then let it run. Requires any Apple Silicon Mac. A great newcomer demo -- follow reaction_diffusion.py as the template for the render loop and _common.py for env setup.