Performance regression: test_conditional_update[v1_indexed] (~517% latest-run slowdown)

Lance SHA: dabd971fe6c8f8da3864ca0b82254b89f4d0cee6\nBenchmark run: https://github.com/lancedb/lance-bench/actions/runs/32656832705\nBenchmark: python/ci_benchmarks/benchmarks/test_merge_insert.py::test_conditional_update[v1_indexed]\nCrate/suite: Python CI benchmarks (merge_insert)\n\nThe regression analysis flags this benchmark at p=0.000750, with the recent four-result window 102.45% slower than older history.\n\nThe latest individual result shows a sharper step:\n- 2c1500f (2026-08-23 10:56 UTC): 8.68 ms\n- dabd971 (2026-08-23 17:57 UTC): 53.54 ms\n\nRecent preceding results were generally 8.4–12.0 ms, so the latest result is approximately 517% slower than the immediately previous run.\n\nThe suspect range is 2c1500f..dabd971. In that range, 30cbafc ("feat(merge_insert): add write_mode to select how merged rows are written", #8423) is directly related and landed shortly before the benchmark, making it the leading candidate.\n\nPlease reproduce the v1 indexed conditional-update benchmark and compare 2c1500f against 30cbafc/dabd971.