Re-evaluate Bokeh in documentation Jupyter notebooks and compare with Plotly

[matz-e]

@kbvw noticed that bokeh<3 is present in the documentation build and that newer versions of Bokeh serialize Neuron objects differently (i.e. fail to do so).

We should see how much work would be involved to switch to plotly, and sort-of ensure that serialization works.

[kbvw]

In Bokeh 3, the serialization to JSON has been reworked: if you want to plot your own container types (such as our hoc.Vector) directly, you need to explicitly add code to serialize them. The documentation on this is not extensive, but from the source, it appears that there are two APIs for this: a Serializable mixin class to explicitly subclass along with a to_serializable method to implement, or a register class method on the Serializer class to add your own encoder function.

We currently only call Bokeh from the notebooks in the documentation, but we advertise there to use it for interactive plotting in the Neuron python workflow. It is currently pinned to bokeh<3, in the main and documentation requirements, as it breaks the notebooks otherwise. This pin is also keeping the numpy<2 pin in place, currently only remaining in the documentation requirements.

I got the feedback that interactive plotting is crucial to demonstrate in the notebooks, be it with Bokeh or with Plotly. That gives the following options:

1. Maintain code with our hoc.Vector to allow Bokeh serialization.
2. Tell users (with the notebooks) to register a simple serialization function when importing Bokeh.
3. Tell users (with the notebooks) to convert hoc.Vectors to lists or np.arrays before plotting.
4. Consider switching to, for example, Plotly for interactive plotting.

I added option 2 in #3061, which removes the pin, and also the remaining Numpy pin. I'd say we can merge that for now, and come back to this if it turns out that registering a serialization function trips up the user too much.

[ramcdougal]

This isn't just a notebook issue.

We give Bokeh code in the example for RangeVarPlot.plot and who knows where else.

The following ought to just work:

import bokeh.plotting as b
bg = b.Figure()
r.plot(bg, line_width=10)
b.show(bg)

So instead of modifying the notebook to tell people to do extra things, is there any reason not to just have __init__.py add the serialization code inside a try/catch block (in case bokeh isn't installed) so that everything continues to just work as in older versions of bokeh?

[kbvw]

I thought about that, and I think that should work and indeed be cleaner, unless the user ends up reloading Bokeh somehow afterwards (or who knows how it gets reloaded), which would lead to a puzzling error.

Not realizing what you said about RangeVarPlot.plot, I thought it would be better to be explicit. But I guess we can import Bokeh upon importing Neuron in case it's installed (and register the function): I suppose it is unlikely to get reloaded after.

EDIT: Changed "implicit" to "explicit"