Guard n_jobs access after conditional column drop

[ethanglaser]

Description

Fixes a crash in --compatibility-mode report generation introduced by #219.

#219 added a block that drops the n_jobs column when it is partially NaN:

if (
    "n_jobs" in results.columns
    and results["n_jobs"].isna().any()
    and results["n_jobs"].notna().any()
):
    results.drop(inplace=True, columns=["n_jobs"])

The pre-existing access below it (added in #184) was left unguarded, so once
the column is dropped, transform_results_to_compatible raises
KeyError: 'n_jobs'.

Three conditions are needed to trigger it, which is why it wasn't caught:

1. --compatibility-mode — the only caller of transform_results_to_compatible
2. n_jobs partially NaN, so the new block drops it. True for any full-suite
   run, since LinearRegression/kNN/ensemble carry n_jobs while Ridge/
   Lasso/ElasticNet do not.
3. More than one environment_name — true when comparing a measurements file
   against a baseline captured on another environment.

Observed on a nightly benchmark report comparing current SPR measurements
against the sklearnex_performance_SPR_2026.1.0 baseline.

---

🤖 Generated with Claude Code