Docs to update for `features/voice-2.0`

Pointers only: where each code change on this branch needs a docs section, and what the section has to cover. Prose to be written by the docs owner.

`docs/content/docs/conversation-simulator-voice-mode/index.mdx`

Setup Voice Mode: new VoiceConfig(record_call=True) option. Records the whole call, off by default.
What A Voice Simulation Produces:
- ConversationalTestCase.call_recording_path: a stereo WAV of the entire call (caller on the left channel, agent on the right), placed by wall-clock time so overlaps and silences are real, not stitched from turn clips. Written to disk as deepeval-call-recording.wav when output_dir is set.
- One INFO log line per conversation from deepeval.simulator: average and worst reply time for the agent and for the caller. Mention it as the way to read latency.
How It Works: on a full-duplex connector (LiveKit rooms, SIP calls) the caller now always listens while it speaks, whether or not the persona interrupts. The speak-then-listen exchange is only used by connectors that cannot carry both voices at once (callback). Consequence worth stating: nothing the agent says while the caller talks or thinks is lost anymore.
FAQs: barge turns made by an interrupting persona count against max_user_simulations.

Turn Detection: the preset window (eager 0.5s / balanced 0.8s / patient 2.5s) is now the ceiling, not the wait. When the running transcript reads as a finished sentence and covers all the speech heard so far, the turn ends after 0.5s of silence. Replies that trail off mid-sentence still wait the full window. Only applies on the duplex path.
Callback / Generic WebSocket (custom connectors): document the supports_duplex property. Returning True makes the simulator run the duplex exchange on that connector even without an interruption policy; it requires stream_uplink, iter_agent_events and stop_uplink.

Who Speaks First: with speaks_first=False on a duplex connector the greeting is collected by the duplex engine (so semantic turn detection applies) and the wait is bounded by hold_timeout, or 15s when unset, after which the caller takes the floor.
Interruptions: note that leaving interruption_behavior unset no longer means half-duplex; the caller still hears everything, it just never cuts in.

How It Works: DuplexExchange accepts policy=None; the barge judge only runs when a policy is set.
What Gets Recorded: Turn.latency_ms is clamped at zero when the agent was already talking before the caller finished (backlogged speech is a wait of zero, not a negative one). Every barge turn counts toward the turn budget.

Verbose logs now print the breakdown as indented JSON with floats rounded to three decimals instead of a Python dict repr. One line under whichever section shows verbose output.

Cross-link the full-call recording and the per-conversation latency log as the two artifacts to inspect when a voice metric scores low.