Fix duplicate weighting in ROD geometric median

[Himerometra]

Preserve sample multiplicities when calculating the geometric median in ROD. Deduplicating the input without retaining counts changes the minimization problem; for example, [0, 0, 0, 10, 20] incorrectly has the same median as [0, 10, 20].

Retain counts in the initialization and modified Weiszfeld updates. Add regression coverage for duplicates, translation, input order and degenerate inputs. Validation: 31 tests passed, including the existing ROD tests. Prepared with AI assistance and pending human review. No issue has been created yet.

[yzhao062]

Thank you for the analysis. On bundled data, weighting the geometric median by duplicate counts lowered ROD accuracy where duplicates are common. On cardio, ROC AUC fell from the 0.935 to 0.953 range to the 0.890 to 0.923 range, and average precision dropped by about 0.10. On shuttle, average precision dropped by 0.06 to 0.23 (3 to 5 seeds, one environment). These runs did not compare against the reference ROD code, so the deduplication may match it. I want to settle the reference-direction question in #726 first, and this pull request stays open until then.