I did few steps to fix the problem.

First of all I tried to run it without stateful flag and with few input vectors instead of one, like this:

Tensors input = tf.keras .Input(shape: new Shape(8, InputVectorSize));

Tensors gruLayer = tf.keras.layers .GRU(units: 32, return_sequences: false) .Apply(input); // <- Error: "Unable to find an entry point named 'TF_GetHandleShapeAndType' in DLL 'tensorflow'."

Tensors output = tf.keras.layers .Dense(SignalOutputVectorSize) .Apply(gruLayer);

model = tf.keras.Model(input, output);

Then I got some new exception: "Unable to find an entry point named 'TF_GetHandleShapeAndType' in DLL 'tensorflow'.". I tried to remove GPU-backend package and install CPU instead. And it work, I can build models with GRU-layers!

Now I have some trouble with fitting, but one more obstackle passed.