[ISSUE-842] Define provenance model for facts, chunks, and evidence

[E2ern1ty]

Purpose of the change

Implements #842 for the geaflow-ai Graph Memory module. Adds the source-lineage contracts needed so extracted facts can point back to the document and chunk they came from, which later citation-aware answers and deletion propagation will build on.

Brief change log

• Add SourceSpan: an immutable span (chunkId, startOffset, endOffset) referencing positions within a chunk, with fail-closed validation.
• Add ProvenanceRef: an immutable lineage record linking an extracted fact to documentId, chunkId, an optional SourceSpan, extractorVersion and schemaVersion. A static validate(ProvenanceRef) fails closed on missing provenance for extracted facts; the optional span must belong to the referenced chunk.
• Provenance records reference source coordinates only and never carry raw document text (guarded by a test asserting serialized provenance contains no raw body).
• Add JSON examples under geaflow-ai/src/test/resources/provenance and JUnit 5 tests covering validation failures, JSON round-trips and equality.

Verifying this change

• Make sure there is a GITHUB_issue field for the change (usually the issue number): [geaflow/ai-memory] Define provenance model for facts, chunks, and evidence #842
• Write adequate unit tests.

mvn -pl geaflow-ai -am test -Dtest='ProvenanceRefTest,SourceSpanTest' passes: 15 tests, 0 failures.

Does this pull request potentially affect one of the following parts

• No. This change is additive: new contracts and tests only, no existing behavior modified.

Documentation

• Contract semantics are documented via Javadoc; no user-facing behavior to document yet.