Remove fairseq2 dependency

[corpaci]

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.

[avidale]

Hi @corpaci !
This package is built on top of fairseq2, and is going to stay this way. The reason is organisational: we, the SONAR maintainers, also have access to the fairseq2 codebase and are in some control about how they co-evolve.

To port SONAR to HF Transformers, on would have to intervene into their codebase, which is unfortunately not easy to organize I tried and failed two years ago; see huggingface/transformers#29646, and, if you like it, you could revive this initiative.

For the particular case of computing text embeddings, I made a port of the SONAR text encoder that does not require touching the Transformers codebase, by fully reusing the NLLB-200 encoder: https://huggingface.co/cointegrated/SONAR_200_text_encoder. So if your goal is only embedding texts, you could use this version.