Performance improvements

[art-from-the-machine]

• For both Piper and XTTS, voice model loading functions are now called as soon as the first NPC is ready, and are run asynchronously while the LLM generates a response
• Added option to Mantella UI to enable / disable lip file generation (or select "Lazy" to skip lip generation for the first returned sentence to improve response times)
• Added tiktoken encoding model locally to data/ folder to avoid connecting to the internet
• Async OpenAI client is created on init to improve response times at the start of a conversation
• Minor performance improvements to conversation summaries loading, NPC searching in character_df, and "trust" calculation
• Added checks to see if context values exist in JSON input before trying to load them
• Added timing logs to many functions for better performance monitoring