plugin: give plugins the error that runner.Run yielded #1883

AfterRunCallback is func(agent.InvocationContext) (plugin.go#L189), so a plugin cannot tell whether runner.Run yielded an error to its caller. The runner hands a root agent's error straight to the caller and keeps running the agent if the caller goes on reading (runner.go#L720-L731). No plugin callback runs for that error.

A plugin that reports how a run ended has to infer it from agent and model callbacks. The BigQuery Agent Analytics plugin does this in #1787, where it writes INVOCATION_COMPLETED or INVOCATION_ERROR. Some runs reach the plugin through exactly the same callbacks whether or not the caller got an error:

• A custom agent that drops its child's error and answers itself, compared with one that forwards the error to a caller that keeps reading. The plugin can't tell these apart, so it has to report one of them wrongly.
• A custom root agent whose Run yields an error and returns, with the caller reading to the end. AfterAgentCallback runs either way (agent.go#L199-L215).
• An agent-level AfterModelCallback that rejects a model response. ADK yields that error without any further plugin callback (base_flow.go#L1046-L1050).

adk-python plugins get the error directly in on_run_error_callback (bigquery_agent_analytics_plugin.py#L9955-L9960).

A fix should let a plugin know, by the end of a run, whether runner.Run yielded an error to its caller and which one, for every runner entry point (Run, RunLive, and the node path).