Add Google Gemini Tokenizer Integration

[ivelin-web]

• Install and import @lenml/tokenizer-gemini.
• If possible, use fromPreTrained() to load the default Gemini encoder.
• Map Gemini models (1.0 Pro, 1.5 Pro, Flash) to this tokenizer. Confirm max context windows (e.g., 32 k for 1.0, up to 1 M for 1.5 Pro).
• Provide a fallback: if the WASM bundle is too large or fails, default to calling Google’s REST countTokens endpoint (requires API key) or use text.length / 4.
• Update tokenizers/index.ts to return a function (text) => tokenizer.encode(text).length when tenant === 'gemini'.
• Write a simple script to feed known‐length text to the Gemini tokenizer and compare with API results for validation.