Add out_balancer for attention_squeeze, similar to nonlin_attention_module.

egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py

@@ -1403,6 +1403,13 @@ class AttentionSqueeze(nn.Module):
                                           prob=_aux_grad_prob_out(),
                                           bias=False, initial_scale=0.05)
 
+        self.out_balancer = ActivationBalancer(
+            channels, channel_dim=-1,
+            min_positive=0.4, max_positive=0.5,
+            min_abs=ScheduledFloat((0.0, 0.002), (8000.0, 0.02), (20000.0, 0.01)),
+        )
+
+
     def forward(self,
                 x: Tensor,
                 attn_weights: Tensor):