Fix tool-call extraction and populate events on the run() path

[ambiorix2099]

Summary

• After run(), AgentResult.tool_calls dropped or misnamed most tool kinds and AgentResult.events was always empty. 13 of the 17 assertions in conductor.ai.agents.testing read one of the two; three pass on the empty list rather than fail.
• Root cause: _extract_tool_calls keyed on a call_ reference-name prefix, which is the provider's tool-call id format and not Conductor's; _SYSTEM_TASK_TYPES skipped five task types that ToolCompiler.TYPE_MAP compiles tool kinds to; survivors were named task_type.lower(). run() never passed events=.
• Fix: identify a tool by its compiled task type, name it from the _agent_tool_name key the dispatch script sets, and derive events from the execution already fetched. Polled task list throughout; the SSE path is untouched.

User impact

One runtime.run() per tree, same agent, a getWeather worker tool and an HTTP tool, Conductor 3.32.0-rc18 on openai/gpt-4o-mini:

	result.tool_calls	result.events
before	[('getweather', {'city': 'Tokyo', '_agent_tool_name': 'getWeather'})]	[]
after	[('getWeather', {'city': 'Tokyo'}), ('serverHealth', …)]	thinking, tool_call, tool_result, tool_call, tool_result, thinking, done

Both runs scheduled the same tasks. assert_no_errors runs on every eval case and was green without evidence; expect_handoff_to, the runner's documented example, failed every time.

Changes

• runtime.py: _is_tool_task / _tool_name / _tool_args replace the prefix check, a SUB_WORKFLOW counting as a tool only with that key, which separates an agent_tool from a handoff; events= at the five AgentResult sites; _task_events shared with both polling streams
• runtime.py: framework agents derive both fields from their execution too, in _run_framework, _run_framework_with_events and both async branches, since the Claude Agent SDK injects a task per tool call
• result.py: one internal-argument key set, so AgentEvent.args and tool_calls agree
• tests/unit/ai/test_tool_extraction.py: new, per tool kind and both provider id formats
• tests/unit/ai/test_runtime.py: test_non_tool_tasks_ignored used a SIMPLE task the server counts as a tool; now INLINE

Fixes #501

[ambiorix2099]

Framework agents: the Claude Agent SDK injects a task per tool call (POST /agent/{id}/tasks) under the Anthropic tool-use id, so those tool calls are in the polled task list and now come back in tool_calls. _run_framework, _run_framework_with_events and both async branches derive tool_calls and events from the execution, as the native path does.

LangChain and LangGraph publish their steps to the event stream rather than as tasks, so their per-step detail still reaches on_event and stream() consumers only. A failed passthrough task shows up either way; an error one of those frameworks handled internally does not.