Speed up common median reference on channel-sized axes

[JESUSROYETH]

Summary

Common median reference currently calls np.median for each trace chunk. For the short channel axis used here, NumPy's full sort path turns out faster than its partition-based median path.

This adds a private reduction helper: one or two numeric channels use their equivalent mean, finite numeric arrays with at least four channels sort and average the middle value(s), and three-channel, NaN, empty, or non-numeric inputs keep np.median. No public API changes.

Benchmark

I ran a complete cold-process run_sorter("tridesclous2"), 20-minute/32-channel recording, n_jobs=1, on an Intel Sapphire Rapids box, five alternating pairs:

	before	after	change
full sorter	170.2 s	150.5 s	−11.6% (19.7 s)

Each of the five paired runs improved, 10.02–18.97%. On SpikeInterface's public 384-channel NHP snippet, the CMR reduction itself comes out 3.89× faster, with identical output hashes.

I ran the common-reference tests too: 16 passed, 1 skipped, Black/style clean. This is complementary to #4564 — it cuts default single-thread work without adding threading.