Fix the per-fold feature selector for genomics data

An AI selected the 50 best of 20,000 genomics features with SelectKBest on the full dataset, then cross-validated a classifier on them. The selector had already seen every test fold's labels, so the cross-validated accuracy of 0.97 was a mirage. The fix reruns the selection inside each fold, on that fold's training samples only.

An assistant wrote select_features_per_fold(feature_rows, labels, fold_ids, top_k). It is in the editor. It has a bug. Find it and fix it.

Input and output:

• feature_rows: one list of feature values per sample.
• labels: each sample's label, at the same index. 1 is positive, anything else is negative.
• fold_ids: each sample's fold id, at the same index. Fold ids are whole numbers, not necessarily 0, 1, 2 and so on.
• top_k: how many features to select in each fold.
• Return one list of selected feature indexes per fold.

Rules the tests check:

• Use only the first n samples, where n is the length of the shortest list. If n is 0, return an empty list.
• The number of features is the length of the first sample's row. A shorter row counts its missing values as 0, and values past that length are ignored.
• Go through the distinct fold ids in ascending order, one output list each.
• A fold's training samples are all the samples with a different fold id. A feature's score is the absolute value of its sum over positive training samples minus its sum over negative training samples.
• Select the top_k features with the highest scores, breaking ties by the lower feature index. Select none if top_k is 0 or less, and all of them if it is more than the number of features.
• Each fold's list is sorted ascending.