docs: improve TTS extension guide with WebSocket patterns and config passthrough

[BenWeekes]

Summary

• Add complete serial WebSocket client and extension examples based on production deepgram_tts implementation
• Add critical WebSocket patterns section: finish_request(), _finalize_request(), TTFB measurement, invalid text handling, cancel/reconnect, audio_start ordering
• Add params passthrough pattern (update_params() convention used by all TTS extensions)
• Add voice-assistant graph wiring section (manifest dependency, property.json graph variant, live testing)

Fixes

• Replace mutable params = {} defaults with Field(default_factory=dict)
• Widen typing from Dict[str, Optional[str]] to dict[str, Any]
• Replace fictional test config filenames with real property_*.json layout
• Add InvalidStatus handling to WebSocket _connect() example
• Fix graph schema to use "graph": {"nodes": ..., "connections": ...}
• Replace misleading "no explicit TTS connections needed" with "copy a known-good graph"

Context

These gaps were identified while building the deepgram_tts extension (TEN-framework/ten-framework#2131). The existing guide covered HTTP mode well but left significant blind spots for WebSocket implementations — particularly around finish_request() (forgetting it hangs the queue), TTFB measurement, invalid text handling for guarder tests, and config params forwarding.

Test plan

• Review rendered MDX for formatting correctness
• Verify code examples compile conceptually against current ten_ai_base API