add ctx.say_filler() for acoustic fillers during long tools

[longcw]

Summary

Adds ctx.say_filler() — an async context manager on RunContext that schedules optional filler speech during a long-running tool call. Fills quiet conversation gaps without burning an LLM round-trip.

@function_tool
async def book_flight(ctx: AsyncRunContext, origin, destination, date) -> str:
    await ctx.update(f"Searching flights from {origin} to {destination}…")
    async with ctx.say_filler("Still searching, hang on a sec.", delay=5):
        await slow_search()
    return result

Design

• Two lanes. ctx.update() is the chat-ctx channel — delivers a real status to the LLM, which voices it. ctx.say_filler() is the acoustic channel — fires session.say() directly, bypasses the LLM.
• Idle-gated. Each fire waits for delay seconds of continuous session-idle before playing, so fillers don't talk over the user or queue behind other speech. State changes (agent_state_changed, user_state_changed) reset the dwell.
• One-shot or repeating. interval=None (default) fires at most once; setting interval makes the scheduler loop with wall-clock cooldown between fires. max_steps caps the number of fires.
• Flexible source. Accepts str or (step: int) -> SpeechHandle | str | None. Callable returning None skips; useful for rotating variants.