Improve: Implement accurate token estimation for cost calculations

[evansantos]

GitSniff finding from PR #49 (Warning level - non-blocking)

Current implementation uses rough approximation: Math.ceil(text.length / 4)

Different embedding models use varying tokenization methods. This approximation may lead to:

• Inaccurate cost tracking for non-English texts
• Misleading budget alerts
• Discrepancies with actual API billing

Suggested improvements:

1. Integrate proper tokenizer (tiktoken for OpenAI models)
2. Add configurable charsPerToken ratio per model
3. Document limitation clearly in config schema

Priority: Medium - impacts cost tracking accuracy

See: https://github.com/openai/tiktoken