Remove language limits: replace spaCy with language-independent tokenization

[azaripov1-oss]

Closes #1

• Removed spaCy dependency — tokenization replaced with a Unicode-aware regex tokenizer that handles CJK character splitting and combining marks, no language-specific configuration required.
• LLM prompts fallback: for languages without dedicated prompts (only en/ru have them), falls back to English prompts with an instruction to generate text in the same language as examples.
• AncSetFit config: requires explicit template for non-en/ru languages; warns when using default templates.
• NER data provider: _parse_record() supports three input formats (tokens+labels, text+spans, bracket format).
• NER evaluation fix: added _predict_from_tokens so evaluation aligns with gold token boundaries.
• Tokenizer tests: offset consistency, no overlap, and full coverage tests on 6 languages (en, ru, zh, ar, hi, fr).