Fix English connector boundary scoring

[aoiminamoto]

Summary

• count English discourse connectors as complete normalized token sequences instead of raw substrings
• preserve repeated connector counts and support future multi-word connectors
• keep Japanese and Chinese connector behavior language-specific
• increment the scoring model version and prevent the trend UI from mixing model versions
• document the corrected metric definition and version behavior

Validation

• Candy is a gift. now reports connectors=0 and score=14.0
• capitalization, punctuation, repeated connectors, and multi-word token sequences have regression coverage
• version-filtered history has regression coverage
• 29 Aoi-maintained application tests pass

The repository-wide legacy test suite could not be collected in the local environment because optional dependencies (fastapi, speech_recognition, and hypothesis) are not installed. The complete aoi_kinabot_app suite, excluding only the FastAPI test that requires the missing fastapi package, passes.

Closes #39