make training false on all build methods

Author: DavidLandup0

keras_hub/src/models/smollm3/smollm3_layers.py

@@ -94,6 +94,7 @@ def build(self, input_shape):
         self.k_proj.build(hidden_states_shape)
         self.v_proj.build(hidden_states_shape)
         self.o_proj.build(hidden_states_shape)
+        self.training = False
         super().build(input_shape)
 
     def call(
@@ -276,6 +277,7 @@ def build(self, input_shape):
             self.intermediate_size,
         )
         self.down_proj.build(down_proj_input_shape)
+        self.training = False
         super().build(input_shape)
 
     def call(self, x):
@@ -397,6 +399,7 @@ def build(self, input_shape):
         self.mlp.build(input_shape)
         self.input_layernorm.build(input_shape)
         self.post_attention_layernorm.build(input_shape)
+        self.training = False
 
         super().build(input_shape)
 
@@ -515,6 +518,7 @@ def build(self, input_shape):
                          - position_ids_shape: (batch_size, seq_len)
         """
         # No internal layers to explicitly build here, as inv_freq is added in __init__
+        self.training = False
         super().build(input_shape)
 
     def call(