Inspect KNN graph during smooth KNN generation and fail if necessary

[viclafargue]

Answers #6490

[csadorf]

@viclafargue Is this ready for review?

[viclafargue]

Yes, it's ready for review.

[divyegala]

@viclafargue looking at the test failures, it looks like the fix is not going to be so easy. Should we slip this to 25.08?

[viclafargue]

@viclafargue looking at the test failures, it looks like the fix is not going to be so easy. Should we slip this to 25.08?

This is not exactly a fix, but rather a safeguard to alert users when a malformed graph is detected. The issue likely does not originate from the PR itself, but from the test setup. Specifically, the test_umap_distance_metrics_fit_transform_trust_on_sparse_input test using the jaccard metric (which compares sets) included samples that had no neighbors with a non-zero distance indicating the presence of many similar sets (n_neighbors occurrences). Increasing the number of features in the test resolves this issue.

[csadorf]

/merge