Add GEMM default Kernel

[crtrott]

The kernel should be able to handle arbitrary layouts as well as scalar types.
It also should have ok performance.

[mhoemmen]

Q: What should users do if the fall-back is host only?

[crtrott]

This kernel works on everything. And the point is that it is a fallback for not having the right type of TPL. It does not work with AD types which use internal parallelization though.

[mhoemmen]

Stokhos deals with this in its Belos adapters by calling GEMM on the underlying entry type (e.g., double). Tpetra may still want to compile those functions for Stokhos, etc. types, though. If we get compilation errors, Tpetra can always work around by wrapping and specializing out for unsupported Scalar types.