perf: true batched inference for TransformerDetector (fixes #23)

[c71qu3]

TransformerDetector.predict_prompt_batch now performs padded batch tokenization and a single model forward pass per configured batch_size, preserving input order and correctly trimming prompt/padding per sample for both tokens and spans. Add strict len(prompts) == len(answers) validation (no silent zip truncation), including in the LLM detector request path.

Tests cover uneven sequence lengths, batch_size 1 and >1, output order, tokens, spans, min_confidence filtering, empty input, and mismatch errors; a spy/stub verifies one forward call per transformer batch without downloading a model.

Summary

Related issue

Type of change

• Bug fix
• Feature
• Documentation
• Tests
• Refactor or maintenance

Testing

• ruff format --check lettucedetect/ lettucedetect_api/ tests/
• ruff check lettucedetect/ lettucedetect_api/ tests/ --extend-exclude lettucedetect/integrations/
• python -m pytest
• Other:

Checklist

• I kept the PR focused on one change.
• I added or updated tests/docs when needed.
• I checked that no secrets, API keys, or credentials are included.

Rights & sign-off (required)

• I certify that I have the right to submit this code and that it may be
  distributed under the repository's MIT license
  (see CONTRIBUTING).