Full support for custom datatypes

[gflegar]

Ginkgo is supposed to be designed to work with user-provided datatypes. However, due to the use of external libraries (cuBLAS and cuSPARSE), which do not have support for this, it is not entirely true.

PR #49 fixes Ginkgo so it can be compiled for such data types, but it doesn't provide alternatives for all external library calls, so parts of Ginkgo will report a non-implemented error for these types.
This issue tracks the progress of these implementations:

• matrix::Dense::compute_dot() (needs custom dot kernel)
• matrix::Dense::apply() (needs custom gemm kernel)
• matrix::Dense::scale() (needs custom scal kernel)
• matrix::Dense::add_scaled() (needs custom axpy kernel)
• matrix::Dense::transpose() (needs custom geam kernel)
• matrix::Dense::conj_transpose() (needs custom geam kernel)
• matrix::Csr::apply() (needs custom spmv kernel)
• matrix::Csr::transpose() (needs custom transpose kernel)
• matrix::Csr::conj_transpose() (needs custom transpose kernel)