Make Leniax simulation fully differentiable

[morgangiraud]

TODO:

• [x] Make leniax pipeline differentiable relative to initial conditions
• [x] Make leniax pipeline differentiable relative to kernel parameters
• [x] Make leniax pipeline differentiable relative to growth function parameters
• [x] Make leniax pipeline differentiable relative to world parameters

Step 1 can be tested with a search over initialization using SGD, a RL-style objectify on the length andsome CPPN/SIREN network.

[morgangiraud]

All right, Leniax is now fully differentiable and I've already been able to learn the Lenia #0 Kernel from noise. Let's see if we can use a neural network for the growth function now.

[morgangiraud]

Let's close this issue now. It's more interesting to replicate Neural CA up to Neural Lenia which generalize the pipeline.