Investigate possible performance regression in `hana::at` for tuples

`hana::at` for tuples seem to be slower than `hana::at_key` for maps on http://metaben.ch. See [this](http://metaben.ch/hetero/clang++-3.9/at/index.html) and [this](http://metaben.ch/hetero/clang++-3.9/at_key/index.html). Is this a genuine bug, or something else? This could have to do with the "improved" support for EBO in tuple introduced in https://github.com/boostorg/hana/pull/313.

Start by comparing the linked metaben.ch benchmarks for hana::at on tuples and hana::at_key on maps. Then inspect the tuple EBO changes in pull request #313 and trace the relevant hana::at entry point. Done means establishing whether the performance difference is a genuine regression or expected behavior, with the cause and next action documented.

Written by the indexing model from the issue text.