feat: add lexical baseline models for ranking tasks

[federetyk]

Addresses #35

Description

This PR adds lexical baseline models to WorkRB for establishing performance bounds on ranking tasks. These baselines complement the existing neural embedding models (BiEncoderModel, JobBERTModel, etc.) by providing lower-bound reference points and enabling future two-stage retrieval pipelines with candidate generation followed by neural re-ranking.

Four models are introduced, all inheriting from ModelInterface and implementing the standard ranking/classification interface. The models accept but ignore ModelInputType parameters, as lexical methods are input-type agnostic. Classification is handled by delegating to ranking, following the same pattern as BiEncoderModel.

The implementations are adapted from the MELO Benchmark repository.

Changes:

• Add BM25Model: BM25 Okapi probabilistic ranking using rank-bm25 library
• Add TfIdfModel: TF-IDF with cosine similarity, supporting word-level or character n-gram tokenization
• Add EditDistanceModel: Levenshtein ratio for string similarity using rapidfuzz library
• Add RandomRankingModel: Random score generation for sanity checking, with optional seed for reproducibility
• Add shared preprocessing: Unicode normalization (NFKD) and configurable lowercasing across all models
• Add rank-bm25 and rapidfuzz dependencies to pyproject.toml
• Add unit tests covering initialization and ranking computation
• Export new models in src/workrb/models/__init__.py

Checklist

• Added new tests for new functionality
• Tested locally with example tasks
• Code follows project style guidelines
• Documentation updated
• No new warnings introduced