[tfjs-layers] fix reshape layer output shape computation

[pvanhaes]

Ensure that output shape is computed correctly when input shape has an unknown batch size.

In Keras, a Reshape layer with batch_input_shape=[None, 16] and target_shape=[-1, 1] has output shape [None, 16, 1] as the reshape operation is being applied per batch.
In TFJS an equivalent Reshape layer would have an output shape [None, None, 1] thus adding an unnecessary unknown dimension. Plus this behavior was not consistent with the python implementation.

---

This change is