Feat/anam add director notes cue support

[sr-anam]

Requires - #6267

What

Adds inline director-note cues to the Anam avatar plugin:
director_note_cue_transform, a tts_text_transforms callable that reads
bracket cues (e.g. [happy], [sad]) in the agent's speech and forwards them to
the Anam engine, keyed by where each lands in the spoken text — so the avatar's
expression changes moment-to-moment, mid-utterance, driven by the LLM's own
output.

Why

The plugin can stream an Anam avatar but gives no way to drive its expression at
runtime. Letting the LLM emit [tag] cues inline means expression follows the
content of what's being said, with no separate control channel and no extra
turn-taking.

Design

• A standard tts_text_transforms entry — composes with other transforms and
  needs no changes to the avatar session.
• Two independent levers cover the real cases: stripped_tags (remove the tag
  before TTS) and forwarded_tags (send it to Anam). So: strip+forward for a TTS
  that would otherwise speak the tag (ElevenLabs), forward-only for a tag both
  face and voice should act on ([laughter]), and a no-op for a TTS that handles
  the tags itself (Cartesia sonic-3.5). Non-preset brackets (e.g. ElevenLabs
  [whispers]) pass through by default.
• Cues ride the existing LiveKit data channel on the director_note_cue topic,
  targeted only at the engine participant (discovered via lk.publish_on_behalf),
  consistent with how the framework's avatar audio output coordinates with its
  worker.
• Offsets are resolved against the TTS-aligned timed transcript, so
  use_tts_aligned_transcript=True is required.

Scoped entirely to livekit-plugins-anam; no change to default behavior for
existing users.