[FEA] Allow cublas wrapper to work on different input/output dtypes

[benfred]

Is your feature request related to a problem? Please describe.
Right now we are constrained to using the same input and output dtypes for all inputs and outputs in the cublas wrapper.
Wrapping the cublas*Ex api's would allow us to partially overcome the restriction and work with mixed dtypes between input and output.

As an example, we're wrapping the cublasSgemm function - which requires all inputs to be a float32, whereas if we changed to wrapping the cublasGemmEx function we could avoid this limitation and have the inputs be a float16 and the output be float32.

Describe the solution you'd like
In detail/cublas_wrapper.cuh we could change:

• cublas<t>gemm to cublasGemmEx
• cublas<T>gemmBatched to cublasgemmBatchedEx
• cublas<t>axpy to cublasaxpyEx
• cublas<t>dot to cublasDotEx
  etc

setting the cudaDataType parameters in the cublas*Ex calls automatically from the template parameters.

Additional context
One issue is that the cublasEx api's don't support all combinations of type parameters. As an example, the cublasGemmEx api requires the A/B input parameters to have the same type afaict - despite taking two different datatype parameters for these values. https://docs.nvidia.com/cuda/cublas/index.html#cublas-GemmEx

Likewise the supported datatype combinations for cublasDotEx requires both inputs and output to have the same type - meaning this change won't have much benefit currently for the linalg::dot function. https://docs.nvidia.com/cuda/cublas/index.html#cublas-dotEx