Fix ComputeACG3D soft merge indexing bug that swaps kept units' data

[JESUSROYETH]

Fixes #4737.

@Arthur031221's issue already has the full diagnosis and a fix in this same shape, this PR just implements it. I re-verified it against current main rather than taking it as given.

SortingAnalyzer.merge_units(..., merging_mode="soft") — the default — silently corrupts the acgs_3d extension for kept units whose index shifts during the merge. ComputeACG3D._merge_extension_data computes old_unit_ids_indices against the merged sorting, then reuses that same index array to read self.data["acgs_3d"] and ["firing_quantiles"] — but those arrays are still laid out in the pre-merge sorting's order. So a kept unit whose index shifted this way comes back holding another unit's 3D-ACG and firing-rate quantiles, with no error and matching shapes, and nothing catches it.

ComputeAutoCorrelograms (same module) and ComputeUnitLocations (unit_locations.py) already do this correctly: they look up self.sorting_analyzer.sorting.id_to_index(unit_id) on the pre-merge sorting for kept units. This fix makes ComputeACG3D do the same thing: it reads the source rows from the sorting the data was actually computed against, and writes them to the destination rows of the merged one.

The regression test (test_acgs_3d_merge) is adapted almost verbatim from the issue, and I extended it to also cover sparse analyzers and a multi-segment recording. It fails on unpatched main and passes with the fix, across merge groups at the start, middle and end of the unit list, and both new_id_strategy values. python -m pytest src/spikeinterface/postprocessing -q and .../core -q still pass otherwise.

Two more things from the issue I left untouched here. test_multi_extensions.py's merge/split check lists 14 extensions and acgs_3d isn't one of them, which is likely why this went unnoticed. Also, the bins value this method returns still reflects the pre-merge unit count after a merge (excluded from the new test's comparison, commented why) — a separate, pre-existing bug in ComputeACG3D.