Fix three bugs in area_of_applicability

[ljwolf]

sorry! Reworking through this code an i caught two bugs in using numpy.percentile (caused by the difference between it and pandas) and one train/test confusion that didn’t affect validity

1. Tukey threshold percentile scale (0.25, 0.75) -> (25, 75) numpy.percentile expects values in [0, 100], not [0, 1]. The old calls were essentially returning the minimum of di_train for both hinges, so IQR collapsed to ~0 and the Tukey cutoff was always the 75th-percentile-ish value of di_train -- then the next line (cutpoint = numpy.maximum(cutpoint, di_train.max())) clamped it up to the max anyway, hiding the bug.

2. Float-threshold percentile scale: numpy.percentile(di_train, t) where t in (0, 1). Same scale issue. Documentation says the user passes a quantile in (0, 1); convert to percentile via threshold * 100.

3. CV-split unpack order: sklearn CrossValidator.split() yields (train_idx, test_idx). The previous loop unpacked them as (test_ix, train_ix), so:

• d_mins[test_ix] = ... was actually d_mins[train_ix_real], overwritten (k-1) times per index (one per fold the point appeared in as training data) instead of filled exactly once for each held-out point.
• hold_to_seen_d was train->test distance, not the intended holdout->in-fold distance.

it’s important to note this is symmetric in Euclidean so the distance values weren't wrong per se, but the rows fed to d_mins[test_ix] = ...min(axis=1) corresponded to the wrong points. Restores the documented semantics: each held-out point's minimum distance to its training fold, recorded once per point.

Further we weren’t using the tukey cutoff, but it’s useful to correct just in case.

[nanophyto]

Good catch thanks for fixing it.