Integrate with River

[martindurant]

https://riverml.xyz/latest/ provides online, per-event machine learning. Whilst we have tried to approach them to accept streamz as a known type of input, it makes sense to just do it here.

The idea would be to have RiverLearn streamz node (either a sink, or a normal node returning a metric) and a RiverPredict node, the pair sharing the same River model or pipeline. The River calls are blocking, but fast - so this can be part of a fast async events system. There are other specific types in River that might emit on a condition, e.g., the drift detector's .change_detected, so we might have a few classes to make, but all of them pretty simple.