The philosophy is that the FlowMaterializer has complete control over
how it interprets the AST, no restrictions. Therefore it only involves
one specified method: materialize() which returns a MaterializedFlow.
Within the ActorBasedFlowMaterializer we materialize Sources and Sinks
that implement the specified SimpleSource/SourceWithKey interfaces (same
for Sinks), others are not supported. These traits are extensible and
they require that an ActorBasedFlowMaterializer is passed into the
factory methods. Other materializers can of course interpret these AST
nodes differently, or they can use the actor-based facilities by
creating a suitable materializer for them to use.
This means that everything is fully extensible, but the infrastructure
we provide concretely for ourselves is built exactly for that and
nothing more. Overgeneralization would just lead nowhere.
Also made FutureSink isActive and implement it using a light-weight
Subscriber instead of a Flow/Transformer.
