checked_mean returns a rounding residue for deltas that cancel on Python < 3.12 because it uses plain sum() #772

dokimi/eval/finite_delta.py:68 checked_mean computes sum(values) / len(values). Before Python 3.12 sum() adds left to right without compensation, so for large values the partial sums round and a set that cancels exactly gives a residue: [1e307] * 6 + [-1e307] * 6 gives -2.079e290 on Python 3.11 (the value in the CI log of #770) and 0.0 on 3.12+ (compensated). The repo compare summary mean (sign_flip_summary) therefore differs by Python version, while the interval uses math.fsum. Reproduced here by running checked_mean with sum replaced by a left-to-right loop (no 3.11 interpreter on the host). Fix: use math.fsum for the finite case (after check_sum_range) and check the tests that compare against sum(...) / n.

Follow-up from the review of #770: https://github.com/intel/dokimi/actions/runs/43636172087/job/132577195767