`accumulator_set` has surprisingly large padding overhead

[k12ish]

I was double checking that accumulator_set is essentially a zero-cost abstraction — But I was surprised to find that's not quite the case!

accumulator_set<double, features<tag::mean, tag::lazy_variance>> acc;
struct HandRolled { std::size_t count; double sum; double sum_sq; };
static_assert(sizeof(HandRolled) == 24);
static_assert(sizeof(acc) == 40); // 66% overhead!!!

I think the overhead comes from the fusion cons list fusion::cons<A, cons<B, cons<C, nil_>>>. Each of the child elements is a full object, so the compiler can't optimize out empty types:

// boost/fusion/container/list/cons.hpp:142-143
car_type car;
cdr_type cdr;

impl	stores	cost after padding in cons
count_impl	size_t	8
sum_impl	double	8
moment<2>_impl	double	8
mean_impl		8
lazy_variance_impl		8

I honestly hoped I could attempt a two line fix in cons.hpp:

- car_type car;
- cdr_type cdr;
+ BOOST_ATTRIBUTE_NO_UNIQUE_ADDRESS car_type car;
+ BOOST_ATTRIBUTE_NO_UNIQUE_ADDRESS cdr_type cdr;

I hotpatched /usr/include/boost/fusion/container/list/cons.hpp, but I couldn't observe any actual memory savings with this on clang or gcc.

Environment

• Boost 1.89.0
• GCC 15.2 / Clang 19
• x86-64 Linux (Arch)

Is this problem worth solving? If I spent longer on this to try and filter out empty types eg. mean_impl lazy_variance_impl at compile time, would this PR be accepted?