Use numpy sort-and-slice for per-cluster statistics

[sharifhsn]

What does this implement/fix?

Replaces per-cluster pandas boolean indexing in _update_derived_values and test_outliers with numpy sort-and-slice.

The existing code loops over every SOM node and selects cells with df[df["clustering"] == cl]. For a 10x10 grid that's 100 full-DataFrame scans per call, each constructing temporary Series and DataFrame objects. In test_outliers the same pattern appears four times per node.

The new code sorts the data once by cluster label, finds contiguous boundaries with np.searchsorted, and slices directly into the numpy array:

sort_idx = np.argsort(labels, kind="stable")
X_sorted = X[sort_idx]
boundaries = np.searchsorted(labels_sorted, np.arange(n_nodes + 1))

for cl in range(n_nodes):
    chunk = X_sorted[boundaries[cl]:boundaries[cl + 1]]

This is fewer lines, avoids the pandas-to-numpy round-trips, and computes statistics on contiguous memory.

Benchmarks

Test FCS file (19,225 cells, 7 markers, 10x10 grid), isolated _update_derived_values (n=50):

	Mean	σ
Before	95.2 ms	2.3 ms
After	50.9 ms	0.8 ms

Outputs are numerically identical (verified: shapes, values, and outlier counts match across all 100 nodes).

No new dependencies. No API changes. All 38 existing tests pass.