Some algorithms don't work with long strings

[Morwenn]

A new tests uncovered issues when sing some sorters with long strings (too big for SSO to apply). I thought it was mostly due to self-move (as with issue #141), but apparently more algorithms break, so while self-moves are a problem, there are other issues at play. Here are some failing sorters:

• block_sorter (with a dynamic buffer following the half policy)
• drop_merge_sorter (see also Strange benchmark results with std::string adrian17/cpp-drop-merge-sort#4)
• grail_sorter (with a dynamic buffer following the sqrt policy)

[Morwenn]

Apparently at least part of the issue with block_sort is due to the fact that I changed a few std::copy to std::move blindly when adapting the algorithms and making sure that it would work with move-only types, which actually introduced a few self-moves.

Specifically I just found such an issue with a call to merge_move where, when merging [first, middle, last), the remaining value of [middle, last) sometimes overrode themselves when they didn't change position. Changing the merge_move call by a buffered_inplace_merge solved this specific issue (see commit f9a0c62), and I'm pretty sure that the remaining issue in block_sort, but also in grail_sort, are similar to this one. After all these algorithms are the only two to use merge_move in the library (occasionally doing it right).