feat: add run_with_events() for a direct event stream

[sjwan01]

Summary

Adds AgentRunner.run_with_events() — a streaming entry point that yields normalized runtime events directly to a bare consumer, with no extension required.

run_stream() forwards only chunks yielded by extensions implementing on_agent_runner_event_stream; a bare consumer receives only run_end. run_with_events() solves this by internally installing a private bridge extension (_EventBridge) that captures chain-only tool events and forwards them alongside streamed tokens.

Event shapes

type	payload
token	{"type": "token", "chunk": str} — one streamed text chunk
tool_call	{"type": "tool_call", "tool_name": str, "tool_call_id": str, "args": dict}
tool_result	{"type": "tool_result", "tool_name": str, "tool_call_id": str, "content": Any, "is_error": bool}
run_end	{"type": "run_end", "session_id": str, "output": str, "new_messages": list, "usage": ...}

Ordering: tool events precede the token chunk of the text that follows them; the stream always ends with run_end. User extensions keep observing events under run_with_events().

Implementation

• run_stream() body extracted into a private _run_stream_impl(streamers) shared by both entry points (no code duplication; the thinking warning is inherited).
• _EventBridge (_internals.py): chain hook queues normalized tool_call/tool_result events; stream hook drains the queue (preserving order) then yields token for TOKEN_STREAM. Never mutates chain data.
• run_stream() behavior unchanged — existing tests untouched.

Known limitations (documented)

• Inherits the pydantic-ai thinking + streaming + tools defect (post-tool-call text may be lost when thinking_enabled=True; run_end.output not a reliable fallback). Warning emitted once on first iteration; prefer run() when thinking is enabled. See README Known issues.
• Events are interception-point snapshots: tool_call.args / tool_result.content show pre-rewrite values; tool_result.is_error is always False (hooks never set it).

Verification

• make check: ruff ✅ mypy strict ✅ 157 passed, 7 skipped (7 = Postgres, run in CI).
• New tests: bare-consumer event stream (order + shapes), warning inheritance, run_stream() unchanged, extension coexistence, early-abort restore, snapshot semantics.