Add support for the SmolLM3 model.

[kadirnar]

🚀 The feature, motivation and pitch

For TTS model training, memory optimization is needed for high batch size values. I am using the liger_kernel library for this. However, there is no support for SmolLM3. Please add this.

Alternatives

No response

Additional context

No response

[vvvdwbvvv]

I would like to work on this issue