Speed up clustering with adjacency matrix compacting and BLAS sign-flip reduction

[larsoner]

Follow-up to #13731. Speeds up permutation cluster functions further:

• Reduce the connected-components graph to only the significant points
• Skip building cluster index-array lists during the permutation loop when possible
• Compute the sign-flip t-test via a single signs @ X reduction (should be memory safe, avoids "unsetting" the applied sign flips)

Investigated the gains and drafted initial implementation with Claude Opus 4.8. I also had it run benchmarks with py-spy (and just wall clock time) to ensure the gains were meaningful. I've also touch a couple examples that should execute faster as well.

[larsoner]

This is a fairly straightforward speedup after #13731 so I'll mark for merge-when-green