Efficient scaling of large language models with mixture of experts and 3D analog in-memory computing