By: lcnbr <im@lcnbr.ch>
avoid rebuilding scalar sums during tensor-network parsing Retain borrowed scalar summands during classification. Reuse the original all-scalar sum and batch the scalar portion of mixed sums with Atom::add_many. Tensor operands continue to use the existing n-ary network addition. Preserve main evaluator aliasing and factorized coefficients. Add coverage for scalar and mixed sums, incompatible tensor summands, and structure discovery.