Fix scatter3d so it handles updates to positions

[nvaytet]

Like the 2d scatter plots, the 3d scatter plot will now handle updates to the data that have a different number of points.

If this is the case, we have to remove the point cloud and add a new one to the scene.

[jl-wynen]

How long does this take approximately? Is it as fast as updating a 2d plot?

[nvaytet]

It depends on the number of points. I did some basic timings: it's about 0.01s for 100_000 points, and ~0.04s for 500_000 points.
But this is only run if the updated values have a different shape than the existing ones, which I am guessing is not super common.

A more relevant question is probably how long does it take to update the positions every time?
We are now running

        self.geometry.attributes["position"].array = np.array(
            [
                self._data.coords[self._x].values.astype('float32'),
                self._data.coords[self._y].values.astype('float32'),
                self._data.coords[self._z].values.astype('float32'),
            ]
        ).T

on every update, which is potentially quite a large allocation?
Before, we only have one array of floats for the colors, now we have 4 (colors + 3 positions).

We could only update if the coords have changed, but we would have to check something like

if any(not sc.identical(old_coords[dim], self._data.coords[dim]) for dim in "xyz"):

I need to check the timings of such a check, maybe it's fast enough?

[jl-wynen]

Yes, check it before you optimise too much! 0.04s seems fine. I don't think anyone expects 60fps.

Just a guess, but maybe you can make the big allocation slightly cheaper by using np.stack or any of its variants instead of np.array([..]).T.

[nvaytet]

The check can actually be pretty fast (10x or more) compared to setting the position, so I added it in.

[jl-wynen]

Do you have to remove the old cloud?

[nvaytet]

This is done above on L111, but maybe it makes more sense to do it here... and remove of the if self.points is not None

[nvaytet]

@jl-wynen this is now ready for another look.
I dug deeper into how things were updated and realized we were setting the colors twice instead of once on every update. So this should also improved performance on large plots (such as the DREAM instrument view).

The logic is probably not super easy to follow from the diff, so we can do a run-through if you'd like that.
I'm also not super happy with member functions that both set attributes on the class and return something (e.g. here), so maybe we can improve that further?

[nvaytet]

Hold on, there are some issues with the 3d clipping...

[nvaytet]

I need to investigate further. The performance impact of changes here is very noticeable on e.g. the DREAM instrument view...