Chapter 16: Training with GPU fails

Hi to all,

I simply ran the solution of exercise 8 but I got:

```
InvalidArgumentError: [_Derived_] Invalid input_h shape: [1,32,30] [19,32,30]
[[{{node CudnnRNNV3}}]]
[[sequential/gru/StatefulPartitionedCall]] [Op:__inference_train_function_3493]

Function call stack:
train_function -> train_function -> train_function
```

I suppose that is related to the usage of GPU because without GPU everything works.
At this point, I tried to recreate the solution from zero, but I got a similar error:

```
UnknownError: [_Derived_] CUDNN_STATUS_BAD_PARAM
in tensorflow/stream_executor/cuda/cuda_dnn.cc(1430): 'cudnnSetRNNDataDescriptor( data_desc.get(), data_type, layout, max_seq_length, batch_size, data_size, seq_lengths_array, (void*)&padding_fill)'
[[{{node cond_29/then/_0/CudnnRNNV3}}]]
[[model_1/lstm_1/StatefulPartitionedCall]] [Op:__inference_distributed_function_5148492]

Function call stack:
distributed_function -> distributed_function -> distributed_function
```

In this case, removing `keras.layers.Masking()(x)` that I added everything works.
The full code of the second experiment is the following:

```
input = keras.layers.Input(shape=[max_length])
x = keras.layers.Reshape((-1, 1))(input)

#x = keras.layers.Masking()(x)
x = keras.layers.LSTM(15, return_sequences=True, activation="relu", kernel_initializer=tf.keras.initializers.GlorotNormal(seed=42))(x)
x = keras.layers.LSTM(5, activation="tanh", kernel_initializer=tf.keras.initializers.GlorotNormal(seed=42))(x)
out = keras.layers.Dense(2, activation="softmax", kernel_initializer=tf.keras.initializers.GlorotNormal(seed=42))(x)

model = keras.Model(inputs=[input], outputs=[out])
```

The version of TF is 2.2.0 by default, but I tried also with 2.1.0.

Do you have any idea of why these errors happen?
Thank you