[API] Move the zero_infinity parameter of CTCLoss from the forward function to the constructor (#76156)

* Enhance CTC loss with zero_infinity parameter

Added 'zero_infinity' parameter to CTC loss initialization and forward method.

* Refactor constructor in loss.py for clarity

* Change zero_infinity type to optional boolean

* Add Optional import to loss.py

* Change default value of zero_infinity to False

* Remove unnecessary import of Optional from typing

* Update zero_infinity handling in ctc_loss function

Refactor zero_infinity parameter handling in CTC loss.

Author: aztice

python/paddle/nn/layer/loss.py

@@ -1340,10 +1340,16 @@ class CTCLoss(Layer):
     blank: int
     reduction: _ReduceMode
 
-    def __init__(self, blank: int = 0, reduction: _ReduceMode = 'mean') -> None:
+    def __init__(
+        self,
+        blank: int = 0,
+        reduction: _ReduceMode = 'mean',
+        zero_infinity: bool = False,
+    ) -> None:
         super().__init__()
         self.blank = blank
         self.reduction = reduction
+        self.zero_infinity = zero_infinity
 
     def forward(
         self,
@@ -1352,7 +1358,6 @@ def forward(
         input_lengths: Tensor,
         label_lengths: Tensor,
         norm_by_times: bool = False,
-        zero_infinity: bool = False,
     ) -> Tensor:
         return paddle.nn.functional.ctc_loss(
             log_probs,
@@ -1362,7 +1367,7 @@ def forward(
             self.blank,
             self.reduction,
             norm_by_times=norm_by_times,
-            zero_infinity=zero_infinity,
+            zero_infinity=self.zero_infinity,
         )