Fix training over existing lora

When training with Lora, and starting with an existing lora weights, current code produces a model with 0 trainable params and training can't work.
Adding the "is_trainable" param allows the loaded peft to be trained and fixes the bug.

src/axolotl/utils/models.py

@@ -402,6 +402,7 @@ def load_lora(model, cfg):
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
+            is_trainable=True,
             device_map=cfg.device_map,
             # torch_dtype=torch.float16,
         )