feat: Domain config through file, Sandhi improvements, and Kannada numeral supportFeature domain file

[kavyamanohar]

Summary

• Domain config standardization (f42de57): Introduced DomainConfig class with factory methods (legal(), medical(), financial(), technical()) and file-based configuration via DomainConfig.from_file(). Bundled
  domain config files added under src/dicterrors/config/.
• Sandhi alignment scope (eb9a904): Sandhi split/merge rules now apply across all token categories except PUNCT and NUMERAL, improving alignment accuracy for agglutinative Indic text.
• Hyphen normalization & script-agnostic word boundaries (23ad7d5): Punctuation splitting is now smart — word-medial hyphens (e.g., "ice-cream") are preserved as single tokens during tokenization. Danda (।) added to punctuation list. Word boundary regex made Unicode/script-agnostic so domain terms match correctly adjacent to Indic characters.
• Kannada numeral normalization (24c69d2): Extended numeral pattern support to recognize Kannada digit sequences alongside existing Malayalam/Hindi patterns. Expanded legal_terms.txt with additional
  terminology.
• use_sandhi toggle + docs update (502b967): Added use_sandhi: bool = True parameter to align_arrays, text_error_rates, token_error_rates, and compute_sample_errors — allowing Sandhi detection to be disabled
  for non-agglutinative languages. Sandhi Detection toggle added to visualizer sidebar. Mismatch penalty multiplier tuned from 0.5 → 0.2. README and CLAUDE.md updated to reflect new API params.

Test plan

• uv run python -c "from dicterrors import text_error_rates; print('OK')" — core import
• uv run python examples/error_report.py — single-sample report
• uv run python examples/batch_evaluate.py — batch evaluation (needs a predictions JSONL)
• streamlit run visualizer.py — verify Sandhi Detection toggle appears and works
• Test use_sandhi=False suppresses Sandhi hits in output