Differentiable indexing with float tensors

Hi,

You can find an example here: Gradient zero for one parameter in custom model resulting in no update