Speed up Wobble peak search with a linear maximum filter

[JESUSROYETH]

Wobble calls argrelmax with order=num_samples - 1 during chunked matching. When candidates remain, SciPy rescans the objective for each offset. So this uses two directional maximum_filter1d passes while preserving strict maxima; arrays with NaN keep the original path.

Five measured alternating baseline/candidate pairs on generated recordings; the VM harness discarded one setup run:

Public path	Fixture	Result
Matching	384 channels, 192 templates	17.932 → 15.581 s (13.11%)
Cold Lupin sort	180 s, 64 channels, 20 units; Intel Cascade Lake VM	120.448 → 106.845 s (11.29%)

The baseline spent 14.76% of matching time in 308 argrelmax calls.

The matching pairs were stable (+11.01–14.45%) and returned the same 28,043 spikes byte for byte. The full-sort pairs were also stable (+10.48–13.11%).

Final-code validation: 700 differential cases; ten byte-identical fixed-template runs; 9 focused tests; the public Wobble test; and formatting. Diff: 88 additions, 5 deletions.

The cold-sort result is corroborating, not equal-workload, because clustering is non-deterministic. The 384-channel full sort remains unvalidated: both arms failed during clustering, before matching, with an unknown singular-matrix cause. No captured recording was benchmarked.

[yger]

Looks great and promising ! Actually, the same argrelmax() should be also in place in circus.py for the OMP peeler, used in a similar manner. Maybe worth also optimizing to be consistent?

[JESUSROYETH]

Thanks @yger, good point on consistency... checking circus.py, the OMP loop doesn't use argrelmax, the peak selection there is already linear with maximum_filter:

result = ndimage.maximum_filter(products[0], size=self.vicinity, mode="constant", cval=0)
cond_2 = np.abs(products[0] - result) < 1e-9

The only argrelmax call in src/ is the one in wobble.py.

One detail: the criteria are not the same. On a flat plateau Circus can keep several tied samples, while this helper keeps the strict argrelmax behaviour exactly, so reusing it in Circus would change its results. The idea would be to keep this PR only for Wobble, and if aligning both makes sense, do it in a separate PR. What do you think?

[yger]

Ok my bad, I wasn't in front of a computer. I'll check the other matching engines, but yes, let's only change if they use argrelmax. I'll read everything more carefully later, thanks a lot !