Investigate the performance of cartesian_product

In a [recent discussion](https://gitter.im/mpark/variant?at=5716c55225b488663625639b) with @mpark, we found out that his implementation of multi visitation on variants was **much** faster than an equivalent implementation based on Hana (in terms of compile-time, we didn't check the runtime). This might hint at a performance problem in `hana::cartesian_product`, or even at a more general performance problem with constexpr-based computation of indices. This has to be looked into.

Start with hana::cartesian_product and the linked Gitter discussion, then compare its compile-time behavior with the cited multi-visitation implementation. Investigate whether the slowdown is specific to cartesian_product or reflects a broader issue with constexpr-based index computation, and document a reproducible comparison and the findings.

Written by the indexing model from the issue text.