Use Unsloth's prepare_model_for_kbit_training for consistency

Changed from peft.prepare_model_for_kbit_training to
unsloth.models._utils.prepare_model_for_kbit_training.

Unsloth's version provides:
- Float32 mixed precision upcasting for LoRA layers
- Better numerical stability
- Consistency with rest of Unsloth codebase

unsloth/models/sentence_transformer.py

@@ -1157,7 +1157,7 @@ class FastSentenceTransformer(FastModel):
 
                 # Prepare for k-bit training if quantized
                 if is_quantized:
-                    from peft import prepare_model_for_kbit_training
+                    from unsloth.models._utils import prepare_model_for_kbit_training
 
                     _gc_for_kbit = (
                         use_gradient_checkpointing