fix: use linear rope_type instead of removed default for transformers compat

The "default" rope_type was removed in recent transformers versions. Using "linear" (with no scaling factor) is functionally equivalent.

modeling_llada2_moe.py

@@ -100,7 +100,7 @@ class LLaDA2MoeRotaryEmbedding(nn.Module):
                 "rope_type", config.rope_scaling.get("type")
             )
         else:
-            self.rope_type = "default"
+            self.rope_type = "linear"
         self.max_seq_len_cached = config.max_position_embeddings
         self.original_max_seq_len = config.max_position_embeddings