Move Buffer::overloaded dispatch helper to top level

[wence-]

GCC versions prior to v12.0 had a bug in class template argument deduction rules: they were incorrectly required to live at namespace scope.

For example, the following definition

struct Foo {
  template <typename... Ts> struct overloaded : Ts... {
    using Ts::operator()...;
  };
  template <typename... Ts>
  overloaded(Ts...) -> overloaded<Ts...>;
};

Produces the compile error

error: deduction guide 'Foo::overloaded(Ts ...) -> Foo::overloaded<Ts ...>' must be declared at namespace scope
    6 |   overloaded(Ts...) -> overloaded<Ts...>;
      |   ^~~~~~~~~~

In contrast, both clang, and later versions of GCC, correctly accept this definition. See
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79501 and https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100983 for details.

Since we still advertise support for GCC 11, and pip-based devcontainer builds use GCC 11, work around this bug by moving the overloaded dispatch helper (used in methods that visit the variant storage) to a top-level anonymous namespace.

[wence-]

Noticed as part of rapidsai/devcontainers#497

[pentschev]

I don't oppose this change, but IIUC this is only required for CUDA 11.x which we explicitly don't support. See conditions where gcc 11.x is required in cuDF versus the only condition rapidsmpf satisfies. To me it seems that the correct solution is to not try to compile rapidsmpf with CUDA 11.x in devcontainers.

[nirandaperera]

LGTM

[wence-]

I don't oppose this change, but IIUC this is only required for CUDA 11.x which we explicitly don't support. See conditions where gcc 11.x is required in cuDF versus the only condition rapidsmpf satisfies. To me it seems that the correct solution is to not try to compile rapidsmpf with CUDA 11.x in devcontainers.

This is required for the pip devcontainer which has GCC 11 as a host. The nvcc host compiler requirement for cuda 12.8 is GCC >= 10; cudf needs GCC >= 11.4.

[wence-]

/merge