Use custom (de)serialization for points/selection

[manzt]

Changes points and selection to use different serde (rather than traitlets List). Currently assumes that points and selection will be 2D numpy arrays.

TODO:

• implement custom JS deserialization
• implement custom JS serialization

NOTE: regl-scatterplot requires data to be in a nested/interleaved structure: [[x0, y0, ...], [x1, y1, ...]]. I'm not sure how feasible it would be, but we could likely improve the overhead of data-transfer further if we could pass columnar data to scatterplot.draw(), e.g.:

scatterplot.draw({
    x: new Float32Array(...),
    y: new Float32Array(...),
    category: new Float32Array(...),
    value: new Float32Array(...),
});

In python, the underlying points/selection arrays could be column-major order (order="F"), which means that we could de-serialize the memoryview on the client into individual columnar typed arrays.

[flekschas]

Looks and works well! Thanks a lot 🎉