test: speed up two slow tests - #34269

- test_calibrated_classifier_cv_works_with_large_confidence_scores: reduce
  n from 1000 to 200. The SGDClassifier never converges on the deliberately
  unscaled data, so the per-fit cost scales with n. A few hundred samples
  still reproduce the large confidence scores from issue 26766 across all
  seeds, dropping the worst-seed runtime from ~18s to ~1s.
- test_cross_validate_params_none: run permutation_test_score with
  n_permutations=5 instead of the default 100; the test only exercises the
  params=None code path, so the permutation count is irrelevant.