perf - coalesce_raw

[cscheid]

With some A/B testing, we find evidence that the code in the PR is faster than that of main (I can share the script later)

$ N_TRIALS=3000 QUARTO_NO_TYPECHECK=1 ./run_experiment_2.py coalesce-raw.csv main tmp-2025-01-21-coalesce-raw quarto render --quiet docs/references/formats/presentations/revealjs.qmd
┏━━━━━━━━━┳━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━┓
┃ Variant ┃ Count ┃ Mean    ┃ Std Dev   ┃
┡━━━━━━━━━╇━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━┩
│ A       │ 10    │ 2.51012 │ 0.0106346 │
│ B       │ 10    │ 2.49222 │ 0.0101653 │
└─────────┴───────┴─────────┴───────────┘
{'outcome': 'significant', 'winner': 'B', 'p_value': np.float64(0.0018407964064409569), 'total_samples': 20, 'alpha_spent': 0.05, 'statistics': {'counts': {'A': 10, 'B': 10}, 'means': {'A': 2.510115218162537, 'B': 2.4922211170196533}, 'variances': {'A': 0.00011309452203267512, 'B': 0.00010333396089645674}}}
Statistical Summary:
===================
Variant     Count         Mean      Std Dev
-------- -------- ------------ ------------
A              10     2.510115     0.010635
B              10     2.492221     0.010165

I used docs/references/formats/presentations/revealjs.qmd from quarto-web because I measured that file to be the one to take the longest time on coalesce_raw in a site render.