Preserve OpenAI token text chunks

[Mirochill]

Summary

• Decode OpenAI tokenizer bytes incrementally when building stored text chunks.
• Preserve UTF-8 characters that span multiple BPE tokens instead of decoding each token in isolation.
• Keep one text chunk per token so existing offsets and embedding windows continue to align.

Fixes #42.

Validation

• git diff --check HEAD~1..HEAD

Tests not run locally.