Our non-lazy adaptors

[Philippe-Cholet]

Fixes #791

Itertools::{kmerge, kmerge_by, tuple_combinations} are neither lazy nor eager. But they must be used since not doing so does almost nothing (and it would be suspicious code to me).
I therefore update the message of the must_use attributes and the laziness tests.

Unless we are willing to roughly wrap those in a LazyInit type (see #791 (comment)) that would probably slow down iteration ("Is it initialized?" each time next is called), I don't see an alternative.

[jswrenn]

Good catch. Could you update #937 after this is merged? I'll cut a release today.