Fix main_channel_id donor selection when merged units disagree

[JESUSROYETH]

merge_units() gives the merged unit main_channel_id from whichever unit is listed first in the merge group on any disagreement, not from the one with the most spikes like the comment says: np.argmax(num_spikes_per_unit.values()) runs on a dict_values view NumPy can't iterate (no __getitem__), so it always returns 0.

Minimal proof: 3 units with spike counts 8/5/7, merging ["b", "a"] returns chB (fewer spikes) instead of chA on main. main_channel_id is set by default for every SortingAnalyzer since #4624.

Fix: look the count up by unit id instead of trusting argmax/position. Doesn't touch the neighbouring note about recomputing from templates for analyzers — only makes the existing spike-count fallback do what its own comment says.

Two regression tests (test_sorting_tools.py, test_sortinganalyzer.py, covering both apply_merges_to_sorting and SortingAnalyzer.merge_units()) fail on main, pass patched; 48 tests total, black clean.