Strengthen prompt injection protection in chat_completion

[0xxreacher]

Summary

This PR strengthens the prompt injection protection in chat_completion() by addressing multiple bypass vectors:

1. Unicode evasion - Zero-width characters, full-width characters
2. Role field injection - Previously unvalidated
3. Partial/modified tags - Space insertion, character encoding

Changes

• Add Unicode normalization function to catch evasion attempts
• Implement comprehensive message validation
• Add 20+ test cases covering all known bypass vectors
• Maintain backward compatibility (no API changes)

Vulnerability Addresses

• Zero-width character bypass (U+200B, U+200C, U+200D)
• Role field manipulation (was not validated)
• Partial tag bypass ("[ INST]" with spaces)
• HTML entity encoding

Testing

All 20+ tests pass:

• Unicode normalization tests
• Injection detection tests
• Dialog validation tests
• Edge case handling
• Integration tests

Security Impact

Fixes potential prompt injection vulnerabilities that could bypass the current safety check.

Breaking Changes

None. This is a security hardening with no API changes.

Checklist

• Code follows project style
• Tests added and passing
• No breaking changes
• Backward compatible
• Security implications reviewed
• Documentation updated with docstrings