Removed stray cuda call

A stray cuda call was preventing this model from being used on devices without a GPU or TPU.

The causal mask is allocated to input_ids.device upon return, as it should.

modeling_aria.py

@@ -620,7 +620,7 @@ class AriaForSequenceEmbedding(AriaPreTrainedModel):
 def precompute_causal_mask(max_seq_len: int):
     return torch.tril(
         torch.ones(max_seq_len, max_seq_len, dtype=torch.bool)
-    ).cuda()
+    )
 
 
 def precompute_freqs_cis(