oneMKL only has `dgmm_batch` variant: doesn't map to cublas

[pen-and-papers]

cublas does not have a specific batch variant of dgmm, it only has cublas<t>dgmm(), see https://docs.nvidia.com/cuda/cublas/#id10
However oneMKL only supports the "batch-style" dgmm_batch interface

This is probably one reason why the cublas backend to oneMKL has no implementation for any dgmm functions. I'm not sure how widely used dgmm is, but I'm guessing since it is a core blas algorithm it should probably be supported.

Is there a reason why there is only the batch style variant of dgmm in oneMKL?

Thanks