The documentation for InteractiveImage is in the source code https://github.com/bokeh/datashader/blob/master/datashader/bokeh_ext.py#L24 , including the bit most relevant to your interest:
Bokeh-based interactive image object that updates on pan/zoom events.
Given a Bokeh plot and a callback function, calls the function
whenever the pan or zoom changes the plot’s extent, regenerating
the image dynamically. Works in a Jupyter/IPython notebook cell,
using the existing notebook kernel Python process (not a separate
Bokeh server). Does not yet support usage outside the notebook,
but could be extened to use Bokeh server in that case.
I.e., this class only supports Jupyter notebooks, and will not currently work with Bokeh server. It should clearly be documented somewhere more obvious, but we are currently investigating whether we can make it work well with Bokeh server, at which point it will move from datashader into Bokeh itself. So, it’s in transition right now; it’s fully usable in Jupyter notebooks, but not currently beyond that. It’s not a lot of code in total, so you could consider adapting it to work with Bokeh server yourself, but there are definitely tricky aspects to it (or we would have done so already!).