[conformer] sdpa default to false (#2362)

Author: Mddct

wenet/transformer/attention.py

@@ -230,7 +230,7 @@ def __init__(self,
                  n_feat: int,
                  dropout_rate: float,
                  key_bias: bool = True,
-                 use_sdpa: bool = True):
+                 use_sdpa: bool = False):
         """Construct an RelPositionMultiHeadedAttention object."""
         super().__init__(n_head, n_feat, dropout_rate, key_bias, use_sdpa)
         # linear transformation for positional encoding