Fix test for recent keras 3 change

When dropout rate is zero, dropout will now create no variables (no
rng seeds). We should set a non-zero dropout to test the presence of
these seeds in tests.

keras_nlp/layers/modeling/cached_multi_head_attention_test.py

@@ -28,6 +28,7 @@ def test_layer_behaviors(self):
             init_kwargs={
                 "num_heads": 2,
                 "key_dim": 4,
+                "dropout": 0.1,
             },
             input_data={
                 "query": random.uniform(shape=(2, 4, 6)),

keras_nlp/layers/modeling/f_net_encoder_test.py

@@ -23,11 +23,11 @@ def test_layer_behaviors(self):
             cls=FNetEncoder,
             init_kwargs={
                 "intermediate_dim": 4,
-                "dropout": 0,
                 "activation": "relu",
                 "layer_norm_epsilon": 1e-5,
                 "kernel_initializer": "HeNormal",
                 "bias_initializer": "Zeros",
+                "dropout": 0.1,
             },
             input_data=random.uniform(shape=(2, 4, 6)),
             expected_output_shape=(2, 4, 6),

keras_nlp/layers/modeling/transformer_decoder_test.py

@@ -36,6 +36,7 @@ def test_layer_behaviors(self, normalize_first):
                 "layer_norm_epsilon": 1e-05,
                 "kernel_initializer": "HeNormal",
                 "bias_initializer": "Zeros",
+                "dropout": 0.1,
             },
             input_data=random.uniform(shape=(2, 4, 6)),
             expected_output_shape=(2, 4, 6),
@@ -59,6 +60,7 @@ def test_layer_behaviors_with_cross_attention(self, normalize_first):
                 "layer_norm_epsilon": 1e-05,
                 "kernel_initializer": "HeNormal",
                 "bias_initializer": "Zeros",
+                "dropout": 0.1,
             },
             input_data={
                 "decoder_sequence": random.uniform(shape=(2, 4, 6)),