Fix seemingly broken padding semantics in praxis/convolutions.py and lingvo/conv_layers_with_time_padding.py

I added a few tests passing NaN showing that at least now it seems to work.

I'll be happy to take comments on how to progress and possibly split this up.

PiperOrigin-RevId: 489153651

Author: lingvo-bot

lingvo/core/conv_layers_with_time_padding.py

@@ -338,7 +338,7 @@ def FProp(self, theta, inputs, paddings):
 
       def _ApplyPadding(tensor_in, padding_in):
         padding_expanded = tf.expand_dims(tf.expand_dims(padding_in, -1), -1)
-        return tensor_in * (1.0 - padding_expanded)
+        return py_utils.ApplyPadding(padding_expanded, tensor_in)
 
       # Zeroing out padded inputs.
       inputs = _ApplyPadding(inputs, paddings)