feat: add finish(), the end-of-stream lifecycle hook

[psyinf]

Stack (bottom → top)

1. feat: add GraphContext, a graph-scoped side channel for stages #9 — GraphContext
2. feat: type-checked merge inputs #10 — type-checked merge inputs
3. feat: add finish(), the end-of-stream lifecycle hook #11 — finish() lifecycle hook (this PR)

Summary

A stage acted only when a message reached it, so a stage that accumulates something (statistics, a batch, an open file) had no defined point at which to flush it. The alternatives were both weak: a destructor (order across a graph is implicit and exceptions cannot propagate) or an in-band end-of-stream message that every stage type has to understand and pass on.

MessageFilter::finish() is that point. It is called once after the last message, defaults to a no-op, and produces no message.

DslFilterGraph<Record, Void> graph(text);
for (auto&& record : records) { graph.filter(std::move(record)); }
graph.finish(); // finishes every stage, in run order

Behaviour

• Forwarded everywhere: FilterGraph, DslFilterGraph, JsonFilterGraph, AnyFilterChain, FanoutFilter, JoinFilter and nested graphs pass it to their stages.
• Run order: DslFilterGraph finishes its stages in run order, so a stage is finished after the stages it reads from.
• Every stage is finished even if one throws (detail::FinishScope); the first exception is rethrown once the others have run. One failing stage cannot keep the rest from flushing.
• No output: finish() returns nothing, so nothing is routed downstream — which would need new merge semantics (every other slot a hole). A stage that wants to hand a final result to the application publishes it through the GraphContext. The name leaves the door open for a flush()-style variant later.
• Owner's decision: a graph that is never finished never flushes, and finishing twice finishes every stage twice. Neither is policed.

Breaking change

AnyMessageFilter::finish() is pure virtual, like setContext in #9: a composite that forgot to forward it would otherwise silently fail to flush its inner stages. Custom AnyMessageFilter implementations must implement it. (Say the word if you'd rather have a no-op default here — it is a one-line change.)

No tick()

Deliberately left out. Time is domain-specific (event time, wall clock, a sensor clock), so a generic signature would fit few users. The README documents the in-band tick message instead: a Tick alternative in the graph's input type, fed through the graph like any other message.

Testing

New LifecycleTests (6 cases): run order in a DSL graph, a nested graph forwarding to its own stages, the compile-time FilterGraph, a JsonFilterGraph with fanout branches and join paths, every stage finishing when one throws, and finishing twice. Full suite: 83/83 locally (MSVC debug, ASan).