Sample, estimate, aggregate: A recipe for causal discovery foundation models