Fix config initialization, serialization, pipeline compatibility, and Apple MPS GPU compatibility

[gokulrejith]

This PR addresses several functional bugs that crash the model during standard inference and loading workflows:

1. Hugging Face pipeline compatibility: Standard HF pipelines pass input_ids and attention_mask and do not provide labels during inference. Added fallback mapping for these arguments and graceful handling of labels=None by returning a SequenceClassifierOutput containing logits, which prevents the pipeline from crashing.
2. Config Initialization & Serialization Fix: Saving and loading the model would crash because FastFitConfig strictly asserted an encoder config was provided. Made encoder optional to comply with Hugging Face's zero-argument config instantiation rules.
3. MPS / Multi-Device Support in SupConLoss: Replaced the hardcoded cuda vs cpu fallback with dynamic device detection (features.device), preventing a crash on Apple Silicon (MPS) and other architectures.
4. Precision Mismatch Fix: Cast labels to FloatTensor instead of DoubleTensor during the fallback loss calculation and explicitly sent them to self.device to prevent precision and device mismatch runtime errors on GPUs.