Temporal Pooling using synapses not cells

[floybix]

Based on the idea from Numenta of minibursts from predicted cells causing extended depolarisation via metabotropic receptors.

• No more temporal pooling excitation. Instead, synapses from predicted cells excite their target cells over multiple time steps.
• No more union pooling; activation level is constant.
• First level layers run exactly the same algorithm as higher layers:
  • Proximal synapses grow to any active synapses, not just stable ones.
  • Winner cells remain the same in continuing active columns unless reset; we may rely on an external timing signal to distinguish repeats.
  • Learn on winner cells only when they become active (even at first level). But might revisit this to learn auto-associatively for pattern completion.

Testing informally on the second-level-motor demo gives plausible results. The activated higher layer columns include some of those from previous predicted states. Obviously needs a lot more testing and development.

[floybix]

Going to merge this because

• i don't want to deal with diverging branches;
• this change makes the code/model simpler;
• what we had before didn't work anyway.