[Fix-18409][Common] Fix DAG.addEdge accepting an edge that creates a cycle when paths converge

[nkuprins]

Was this PR generated or assisted by AI?

YES - the bug was found using AI-assisted static analysis, and the fix and regression test were AI-assisted. I verified everything myself: I confirmed the failure by running the new test against unmodified dev (fails), reviewed the logic, and confirmed the fix (passes) with the full dolphinscheduler-common suite green.

Purpose of the pull request

Closes #18409. DAG.isLegalAddEdge bounds its cycle-check BFS with a node-count budget (--verticesCount > 0) instead of tracking visited nodes. On graphs where paths converge, a node is enqueued once per incoming edge, exhausting the budget before the BFS reaches fromNode, so a cycle-creating edge is wrongly judged legal and inserted. This replaces the budget with a visited-set - the correct BFS termination condition - which cannot be defeated by path convergence and also avoids redundant re-enqueueing on dense graphs.

Brief change log

• common/graph/DAG.java: track visited nodes in isLegalAddEdge instead of a decrementing verticesCount budget; add HashSet import.
• common/graph/DAGTest.java: add testCycleWithConvergingPaths, a fan-shaped graph asserting the cycle-creating back-edge is rejected and no cycle is introduced.

Verify this pull request

This change added tests and can be verified as follows:

• testCycleWithConvergingPaths fails on current dev (addEdge returns true for a cycle-creating edge) and passes with the fix.
• Full dolphinscheduler-common suite passes (139 tests, 0 failures) on JDK 8.
• Spotless format check passes.

[boring-cyborg]

Thanks for opening this pull request! Please check out our contributing guidelines. (https://github.com/apache/dolphinscheduler/blob/dev/docs/docs/en/contribute/join/pull-request.md)