Refactor BLAS interface for scalars

[jennloe]

It would be helpful to refactor the KK BLAS interface to accept additional options for scalars. This would be useful for the Belos refactor to remove the Teuchos::SerialDenseMatrix.

We need for the BLAS interfaces to work with a scalar value that is on the device. For instance, I would like to be able to call KokkosBlas::scal(y,alpha,x); where alpha is either a rank-1 or rank-0 view on device. Similar comments apply to KokkosBlas::axpy.

It would be preferred that the interfaces be consistent between all the BLAS functions- i.e. they all accept rank-1 views for scalars or they all accept rank-0 views. This is not currently the case. For example, the dot function can return a scalar only to a rank-0 view, while the axpby function only accepts scalars in rank-1 views.

Thanks!