More systematic gcc & clang coverage

[rwgk]

Description

Based on #4074 (comment)

Drops GCC 10 --std=c++2a (because it is an odd choice), adds 6 GCC / -std combinations not covered already but assumed to be what users would pick.

Also adds clang 11, 12, 13 (not covered at all before), with --std=c++20.

Resolves GCC 11 & 12 "redundant move in return statement" warnings.

Suppresses misleading clang warnings in the same locations.

Suggested changelog entry:

[Skylion007]

This looks good to me to me on the CI front, but I was confused why the pragrams are still in the include if the std::move is removed?

[Skylion007]

If you remove the std::move are the pragmas necessary here, anymore?

[rwgk]

Here a GCC 11 & 12 warning is at conflict with a clang 7 through 12 warning.

When GCC 11 & 12 complained, I simply removed the std::move. (Thinking: "Let's see what happens.")
Next CI: many clang failures.
Next CI: complete the clang coverage to better understand what is going on.

Looking at the results I concluded (guessed TBH): clang was wrong for a long time, but that was fixed in clang 13. The fix appears to stick: clang 14 & 15 agree.

Next CI: with pragmas for clang added.

I considered introducing something like

PYBIND11_SUPPRESS_MISLEADING_CLANG_WARNING_CALL_STD_MOVE_EXPLICITLY_BEGIN
...
PYBIND11_SUPPRESS_MISLEADING_CLANG_WARNING_CALL_STD_MOVE_EXPLICITLY_END

or

    return PYBIND11_KNOWN_REDUNDANT_STD_MOVE(result);

but decided against it because there are only two affected locations, and the implementation you see now is easier to understand just looking at the local code, without having to look elsewhere.

[rwgk]

@henryiii I'll merge this now, to have this included in the next smart_holder update. Please let me know any suggestions, I'll work on them in a follow-on PR.

@Skylion007 thanks for reviewing!