Remove fairseq2 dependency

SONAR currently depends on `fairseq2>=0.5.2` for transformer architecture, tokenization, speech models, data loading, sequence generation, and model registry. This limits adoption.

Goal: replace all fairseq2 usage with PyTorch-native implementations and/or HuggingFace transformers, making SONAR installable with standard `pip install` and compatible with a wider range of environments & make SONAR easier to use for embedding-only downstream applications while keeping the larger migration incremental.

Start by locating every fairseq2 usage in SONAR and grouping the listed responsibilities: architecture, tokenization, speech models, data loading, sequence generation, and model registry. Compare PyTorch-native and HuggingFace transformers options, then verify that standard pip installation works and embedding-only downstream use remains supported while the migration stays incremental.

Written by the indexing model from the issue text.