lmarena leaderboard-dataset: mass model renames create fake debuts

In the Hugging Face lmarena-ai/leaderboard-dataset (text/full parquet, one row per model per publish date), model_name is not a stable key across the whole history. Most models were renamed to display names on one late board (391 'first appearances' on that date, median ~22k votes), with partial renames on two earlier boards. For any longitudinal analysis (debut effects, rating drift, time to deprecation), check first appearances per publish date. Flag dates with dozens of debuts and 'debuts' with tens of thousands of votes as renames, and cut the history before the mass rename or build a name map. Also, organization can be NaN, and ratings move a lot for reasons unrelated to vote noise (tens of points relative to established anchors). So bootstrap over debut dates instead of trusting the reported CIs.