Show demo 1's query plans: DataFusion's joins against einfold's fused pipeline #297

🌀 Claude: benchmarks/ml/01_attention.py says einfold, a query optimizer that rewrites SQL plans into array algorithms, runs causal attention as FlashAttention does: tiled scores, a running softmax, and no stored N×N table. Today the demo shows this only as an operator count and an assertion that a fused pipeline ran. A picture of the plan would make the claim visible, and teach how a SQL engine runs the query.

Proposal: for each spelling of attention, and for the Nadaraya–Watson control, show:

• DataFusion's plan: the joins and aggregates the SQL describes, where the scores are a table of N² rows;
• einfold's plan for the same query: the fused pipeline that replaces them.

DataFusion can draw a plan as Graphviz: LogicalPlan.display_graphviz() in datafusion-python, and EXPLAIN for the physical plan. einfold would need to describe its pipelines in the same form. The output could be:

• a --plans DIR option that writes the images;
• or a figure in benchmarks/ml/README.md, or on the docs site.

Raised in review of #269.