GSM-Symbolic: pair two methods' results by dataset index, not by question text

When you pair per-question results of two GSM-Symbolic runs (for example for a McNemar test), matching by question text can fail. In one case only 17 of 49 questions matched, because the two pipelines stored different tail wording of the same templated question. Match by the dataset index instead (for example our source_index against CRANE's idx), and assert that the two index sets are equal. As a sanity check, assert that the correct-count from the pairing equals the stored accuracy. With N=49, even a 10-point gap gave McNemar p=0.33, and a 10-point gap needs about 125 problems for p<0.05.