Speed up cluster index list construction in permutation cluster tests

[Lucakaemmer]

Reference issue (if any)

Follow-up to #13731 and #14057.

What does this implement/fix?

Since #13731/#14057, _get_clusters_st and _get_components build the list of cluster index arrays with

[active[labels == id_] for id_ in np.unique(labels)]

which goes through labels once per cluster. With many clusters this is the main cost of a permutation, especially for TFCE, where it happens at every threshold. This PR replaces it with a stable argsort followed by np.split, in a small helper used by both functions. The result is the same list (same clusters, same order), just built in one pass. A small test for the helper is included.

Timings on a 12 x 143 x 143 lattice (245,388 points, 12 "subjects" of Gaussian noise, tail=1; Apple M3, numpy 2.5.2, scipy 1.18.1, numba 0.67.0):

	main	this PR
_find_clusters, TFCE (start=0, step=0.2), global adjacency	2.03 s	0.50 s
_find_clusters, TFCE, spatial adjacency with max_step=1	1.91 s	0.40 s
_find_clusters, threshold=0.5 (14,788 clusters)	0.28 s	0.03 s
permutation_cluster_1samp_test, TFCE, 20 permutations	57 s	12 s

Cluster indices, cluster sums, TFCE scores, T_obs, H0 and p-values are identical between main and this branch on these inputs.

Additional information

Drafted and tested the initial implementation with Claude Fable 5.

[larsoner]

Makes sense that this would speed up TFCE since it has so many independent "clusters" (treating each point as one potentially). LGTM, marking for merge-when-green and I approved CIs. Thanks in advance @Lucakaemmer !

[welcome]

🎉 Congrats on merging your first pull request! 🥳 Looking forward to seeing more from you in the future! 💪