fix(workflow): persist Agent caller before node execution

[zyz619963502zyz]

Summary

Fixes #39546.

Graphon 0.6.0 runs layer on_node_run_start hooks synchronously in the worker before node.run() can advance. This change uses that boundary for Agent v2 to commit its WorkflowNodeExecutionModel caller row before participant/session setup begins.

• synchronously persist the Agent v2 caller in WorkflowPersistenceLayer.on_node_run_start
• serialize concurrent Agent start writes so node sequence allocation remains consistent
• enrich the already-persisted execution when the normal NodeRunStartedEvent reaches the dispatcher instead of inserting it again
• propagate synchronous persistence failures back into DifyAgentNode, before binding resolution or participant allocation
• remove the old 60-attempt / 3-second caller visibility polling loop

Non-Agent nodes retain the existing event-driven persistence path.

Tests

• pytest api/tests/unit_tests/core/app/workflow/test_persistence_layer.py api/tests/unit_tests/core/workflow/nodes/agent_v2/test_agent_node.py api/tests/unit_tests/core/workflow/nodes/agent_v2/test_session_store.py -q — 66 passed
• Ruff on all six changed files
• py_compile on production modules
• git diff --check

The new tests deterministically cover both the worker-first ordering and persistence failure before participant allocation; no sleeps or timing polls are used.