[ITensors] [ENHANCEMENT] Sparse output for `onehot`

[stuartthomas25]

The onehot function currently outputs a dense tensor which is not very efficient to contract. Ideally, this function would output an ITensor with sparse storage.

It seems like there is not yet an NDTensor TensorStorage type which implements a sparse array though I see that there is movement in overhauling the TensorStorage system (#1250).

[mtfishman]

Yes, we are doing a complete rewrite of NDTensors which will replace the Tensor type and TensorStorage types. The new design will include a general elementwise SparseArrayDOK type (using a dictionary-of-keys format), a new BlockSparseArray type, a new DiagonalArray type, along with better support for array slicing of all of those types. With better support for slicing it would mitigate the need for onehot in the first place, since slicing is the main use case for it. The plan would be to base onehot on SparseArrayDOK or the OneElement type from FillArrays.jl and use slicing under the hood for tensor contraction.