feat: Geometric Turing Machine v4 for ARC-AGI

[Concode0]

Summary

• Adds the Geometric Turing Machine (GTM), a few-shot grid transformation architecture for ARC-AGI tasks
• Uses Projective Geometric Algebra Cl(3,0,1) as the CPU algebra, enabling both rotation and translation via motor sandwich products
• Three-phase training: warmup → circuit search → adaptive computation (PonderNet)
• Key components: GridCodec (PGA encoding), GeometricCPU (motor + ColorUnit), GeometricSuperpositionSearch, RuleAggregator (cross-attention rule memory), TuringVM, GTMAnalyzer
• Run: uv run main.py task=gtm

Test plan

• All 343 existing tests pass
• Forward + backward pass verified (fixed-step and ACT modes)
• Full training run

[Concode0]

Temporarily pivoting GSS development toward theoretical grounding. Current hypothesis exploration is a valuable PoC, but I'm moving toward a more rigid geometric framework to eliminate the current state-stagnation issues. Updates to follow.