chore: sync sdk code with DeepLearning repo

[bgotthold-aai]

v0.64.16

• Streaming connections now retry transient handshake failures. New StreamingClientOptions fields —
  max_connection_retries (default 2) and connection_retry_delay (default 0.5s) — automatically retry timeouts and
  network drops during the WebSocket handshake. HTTP-level rejections (auth, quota, bad request) still fail fast.
  The handshake connect_timeout is now configurable (default 1.0s).
• Graceful disconnect waits for the server's final messages. disconnect(terminate=True) now blocks (up to the new
  terminate_timeout, default 5.0s) for the final Turn and TerminationEvent to dispatch before teardown, so the
  closing transcript and total-audio-duration report are no longer dropped. Applies to both the sync and async
  clients.
• BeginEvent now exposes session configuration. A new SessionConfiguration model (model, mode, api_version) is
  available on BeginEvent.configuration, letting you read back the negotiated streaming settings. Parsing tolerates
  unknown/null mode values so a new server-side mode won't drop the event.
• SyncTranscriptResponse word type unified. words now uses the standard Word model instead of the separate
  SyncWord type (now removed). The inference_time_ms field was also removed from the sync response.
• README no longer carries the "testing and light usage only" warning, reflecting the SDK's support for production
  use.