std::unique_ptr deleter unit tests.

[iwanders]

Description

Hello again @rwgk , I saw your call for unit test help in the pybind#2839 (comment) issue (long time watcher on that issue ;) ), I had some time to spare so I took a stab at your unit test request.

So I wasn't too familiar with the original issue, but from the looks of it the main thing is that we need to be able to handle and test custom deleters that are a closure are called at the correct times, only when the delete actually happens.

I tried to think some unit tests around this and came up with these two:

• from_unique_ptr_with_std_function_deleter+as_lvalue_ref+destructor_called
  • Checks that std::function deleter is not invoked when moved from a unique pointer into the smart holder.
  • Verifies the deleter is invoked when the smart holder is destructed.
• from_unique_ptr_with_std_function_deleter+as_raw_ptr_release_ownership
  • Checks that anything with a std::function deleter cannot be disowned.

Then I saw the unique_ptr -> holder -> unique_ptr roundtrip unit test... and realised that the deleter should also survive that roundtrip, so I added;

• from_unique_ptr_std_function_with_deleter+as_unique_ptr_with_std_function_deleter1
  • Confirms that the std::function deleter survives the roundtrip.

Problem is, that third unit tests failed... because the as_unique_ptr method on the smart holder didn't extract the deleter from the shared pointer to pass it on to the returned unique_ptr. In 1ce7053 and 84033fd I tried to remedy this issue, because I think we should pass along the deleter if we go from a smart holder back to a unique pointer.

Unfortunately, the unit tests with:

    REQUIRE_THROWS_WITH((hld.as_unique_ptr<int, helpers::functor_other_delete<int>>()),
                        "Incompatible unique_ptr deleter (as_unique_ptr).");

Made this a bit problematic, as that made it impossible to just do an assign into the deleter type of the unique pointer, so there's a tiny trait system here to ensure the assign only happens if the type is actually assignable (preventing compilation errors with the hld.as_unique_ptr<int, helpers::functor_other_delete<int>>() unit tests).

Upon writing this, I think we should actually throw if we have a deleter but are using a return type that cannot accomodate our custom deleter.

Either way, this PR is definitely not ready, I expect formatters & linters will complain about many things, but thought I'd file it anyway to open discussion.

[iwanders]

This obviously needs to go... I needed it to debug something.

[iwanders]

Closing based on discussion in pybind#4850