I don't think this is quite right, the index here is the index into the mask, not the index of the index. Here's an example from the curve reverse node:

threading::parallel_for(selection.index_range(), 128, [&](IndexRange range) {
  for (const int i : range) {
    splines[selection[i]]->reverse();
  }
});

Constructing a transform matrix from an Euler rotation is very slow. We should try to avoid this constructor whenever we can.

I wonder if this matrix is actually necessary, or if float4x4::apply_scale(scale) would work.