Lambdas in various cases of std::apply in xfunction.hpp don't return correct types

[vakokako]

After an update from #2876, the following code stopped compiling:

auto values_original = xt::xtensor<double, 1>{ 1., 2., 3. };
auto indexes = xt::arange<size_t>(values_original.size());
auto map = [&](const size_t& index) -> auto& { return values_original[index]; };
auto values_mapped = xt::vectorize(map)(indexes);
for (size_t i = 0; i < values_original.size(); ++i) {
    EXPECT_TRUE(&(values_mapped[i]) == &(values_original[i]));
}

The reason is in this function:

xtensor/include/xtensor/core/xfunction.hpp

Lines 1108 to 1118 in 5caa64d

	template <class F, class... CT>
	inline auto xfunction_stepper<F, CT...>::operator*() const -> reference
	{
	return std::apply(
	[&](auto&... e)
	{
	return (p_f->m_f)(*e...);
	},
	m_st
	);
	}

Here, the lambda needs to have a return type reference instead of returning by value.

The same problem occurs in every use case of std::apply in xfunction.hpp.

[JohanMabille]

Wow good catch! Unbelievable that we haven't catched it earlier!