fix(graph): prevent duplicated append results when merging serial subgraphs

[gaozexun]

Describe what this PR does / why we need it

This PR fixes duplicated/over-amplified AppendStrategy results when parent graph merges serial subgraph outputs (Issue #4515).

Before the fix, serial subgraph executions could produce duplicated output values:

• two distinct subgraphs in serial -> expected 2, got 3
• same subgraph instance reused by two nodes -> expected 2, got 4
• parent node + two serial subgraphs -> expected 3, got 7

Does this pull request fix one issue?

Closes #4515

Describe how you did it

Two changes were applied:

1. Isolate subgraph checkpoint namespace by nodeId

• In SubCompiledGraphNodeAction, once subgraph saver exists, always assign deterministic node-scoped threadId:
  • ${parentThreadId}_${subGraphId(nodeId)} or subGraphId(nodeId)
• This avoids checkpoint/state pollution across different parent nodes using subgraphs.

2. Reset append-based keys before subgraph execution

• Added OverAllState.withResetKeys(Set<String>).
• In SubCompiledGraphNodeAction, build subgraph initial state by removing all keys whose strategy is AppendStrategy.
• This ensures each subgraph run starts append accumulation from a clean baseline instead of inheriting parent historical append state.

Also added a non-mock reproduction test class:

• Issue4515ReproductionTest (3 scenarios)

Describe how to verify it

Run:

./mvnw -pl :spring-ai-alibaba-graph-core -Dtest=Issue4515ReproductionTest test
./mvnw -pl :spring-ai-alibaba-graph-core -Dtest=Issue4515ReproductionTest,CompiledSubGraphTest,SubGraphTest,StateGraphTest test