Add example: Smart Turn detection - #92

Adds docs/examples/smart_turn_detection/ — using an open end-of-turn model as
the turn detector in an AgentSession.

turn_handling["turn_detection"] accepts any object implementing the streaming
turn-detector protocol, but there is no example of supplying one. This shows the
shape:

proc.userdata["turn_detector"] = SmartTurnDetector(
    model="smart-turn-v3",
    on_prediction=log_prediction,
)

turn_handling={
    "turn_detection": ctx.proc.userdata["turn_detector"],
    "endpointing": {"min_delay": 0.3, "max_delay": 2.5},
}

on_prediction prints each decision, so the model is visible without a
frontend:

end of turn 1.2%  -> incomplete (threshold 0.50, 55 ms)
end of turn 94.5% -> complete   (threshold 0.50, 58 ms)

STT, LLM and TTS are the same inference.* models as
docs/examples/metrics_vad/, and VAD is the same Silero plugin. The one new
dependency is smart-turn-livekit, which loads the ONNX graph and adapts it to
the protocol — a few MB at int8, CPU-only, fetched from Hugging Face on first
use.

Default weights are upstream Smart Turn v3.
model= selects a different graph and model_path= loads your own.

Plugin: https://github.com/santhosh-005/smart-turn-livekit (BSD-2-Clause)

Verified on livekit-agents 1.7.0: frontmatter parses with
docs/tools/generate_index.py, the session builds with the detector attached,
and real speech through the detector returns predictions in 55–123 ms on one CPU
thread. Not tested with a live call.