Speed up collision-aware amplitude scaling

[JESUSROYETH]

Collision-aware amplitude scaling scans the whole chunk again for every selected spike, just to recover a position that's already known once the non-margin subset is built. So this bookkeeping step ends up quadratic in the number of spikes per chunk.

Changes

The fix keeps those flat indices and passes them to find_collisions(), building each collision group once instead of concatenating it over and over. The existing helper call still works as is, the fitting arithmetic doesn't change.

Validation

I ran two seven-pair fresh-process sweeps after one warm-up, on 100-unit, 64-channel generated recordings, through the public analyzer.compute("amplitude_scalings", handle_collisions=True) path with one worker and 1 s chunks. Times below are complete extension calls: 20 s at 4 Hz and 10 s at 20 Hz.

CPU VM	~400 spikes/chunk	~1,995 spikes/chunk
Intel Ice Lake	1.826 → 1.552 s (-15.0%)	11.604 → 9.945 s (-14.3%)
AMD Milan	1.326 → 1.118 s (-15.7%)	8.172 → 6.992 s (-14.4%)
ARM Neoverse N1	1.595 → 1.350 s (-15.3%)	10.239 → 8.770 s (-14.4%)

At high density this saves 118-166 ms per chunk, and 17,375 of the 19,954 spikes need a collision fit. All before/after ranges are separated, and all 42 measured pairs came back with identical arrays and collision counts. The public MEARec fixture also matched exactly. The complete module passes its 6 tests across sparse/dense analyzers, two segments, and memory/binary/zarr formats; Black is clean.