Adding support for LightOnOCR_2_1B

Hey I am trying to contribute my implementation of https://huggingface.co/lightonai/LightOnOCR-2-1B to candle. However I am running into bit of a issue where the model has a dependency on qwen3.

The weights for LightOnOCR are in the format:
```model.language_model.embed_tokens.weight```

But qwen3 expects
```model.embed_tokens.weight```

This wouldn't have been much of a problem if not for the implementation of qwen3 baking in the ```model``` part. This is my first time doing open source so I am not sure if I am allowed to change the implementation for qwen3 so any help would be greatly appreciated.

Start by reading Candle's qwen3 implementation and comparing its expected model.embed_tokens.weight names with LightOnOCR's model.language_model.embed_tokens.weight format. Determine the appropriate model-loading or qwen3 integration point, then verify that the LightOnOCR-2-1B weights can be loaded without breaking existing qwen3 support.

Written by the indexing model from the issue text.