Implement a novel attention mechanism where each head corresponds to
a rooted tree from OEIS A000081 (number of unlabeled rooted trees).

- rooted_trees.py: Generate rooted trees with Butcher coefficients
  (order, symmetry σ, density γ, α coefficient)
- attention_heads.py: Map trees to attention head specifications
- elementary_differentials.py: Tensor contraction patterns from
  Runge-Kutta theory's elementary differentials
- hyper_attention.py: Full transformer with tree-structured attention
  - ATen(n) layer has a(n) heads: 1,1,2,4,9,20,48,...
  - Tree topology induces attention patterns (sequential/parallel)
  - Optional PyTorch implementation with TreeHead modules
- demo.py: Comprehensive demonstration of the architecture

The architecture encodes combinatorial structure directly into
attention, with each tree's properties determining tensor dimensions.

Claude/attention tensor trees mu nj d