Mark curation points with a single vstack instead of a loop

[aymuos15]

Description

What is this PR — [x] Bug fix

CurationWidget.mark_point_as_type appended selected points one at a time, reassigning the Points.data property inside a loop over selected_data. Each assignment re-validates the whole array and emits change events, so marking N points is O(N²).

This PR indexes all selected rows at once and does a single vstack/.data assignment. Same rows in the same order.

selected_points = layer.data[list(layer.selected_data)]
destination_layer.data = np.vstack((destination_layer.data, selected_points))

Benchmark

Time — old loop vs. single vstack against a real napari.layers.Points layer (destination layer = 500 points):

selected points	loop (current)	single vstack	speedup
10	21 ms	5.7 ms	4×
25	45 ms	5.8 ms	8×
50	90 ms	6.0 ms	15×
100	176 ms	6.2 ms	29×
200	359 ms	6.2 ms	57×
500	916 ms	6.2 ms	148×
1000	1.9 s	6.6 ms	292×
2000	4.5 s	7.6 ms	592×
5000	15.1 s	9.4 ms	1613×

Peak memory — numpy allocations (tracemalloc). The single version also lowers peak memory ~2×, because the loop rebuilds the whole growing destination each iteration (holding two copies at its peak) while the single vstack touches it once:

selected points	loop (current)	single vstack
50	4.80 MB	2.40 MB
500	4.82 MB	2.42 MB
5000	5.04 MB	2.64 MB

System: Intel i7-12800H (20 threads), 30 GiB RAM, Ubuntu 22.04 (kernel 6.8.0-111), Python 3.11.15, numpy 2.4.2, napari 0.6.6.

How has this PR been tested?

KERAS_BACKEND=torch pytest tests/napari/test_curation.py passes (existing tests cover mark_point_as_type; behaviour unchanged). Benchmark outputs verified identical (same shape and values).

Is this a breaking change?

No.

Checklist

• The code has been tested locally
• Tests added — behaviour unchanged, already covered by test_curation.py
• Docs updated — n/a