Support compiling C++ modules with Clang >= 16

[chrisnovakovic]

The flags necessary for Clang to compile C++ modules changed between Clang 15 and 17:

• Clang >= 16 (and therefore Apple Clang >= 15, which are built from LLVM >= 16) just require C++20 features to be enabled (-std=c++20) in order to enable modules, while older versions of Clang explicitly require modules to be enabled via the -fmodules-ts option. Choose the appropriate flag based on the detected Clang version.
• Clang >= 17 (and Apple Clang >= 16) require a dependent module's name to be specified along with its path with -fmodule-file, otherwise a warning (-Weager-load-cxx-named-modules) is outputted, but specifying the module's name at all is a syntax error with Clang < 16 (and Apple Clang < 15). When using Clang >= 16, specify the basename of the module file as the module name; nothing in the C++ specification says that a single module has to be declared in a file of the same name, but it'll be the case most of the time (and will fail loudly in the other cases).

Fixes #27.

[chrisnovakovic]

Of course Apple Clang has to behave differently to mainline Clang in this specific instance...

[peterebden]

Makes sense. Was ages ago that I looked at the module stuff though; I'm not surprised it doesn't match up any more, but I also hear they've been tricky to support in general - I don't know to what degree we have something that works in the real world any more. (Although it is noted as experimental in the comment at least!)

[chrisnovakovic]

Makes sense. Was ages ago that I looked at the module stuff though; I'm not surprised it doesn't match up any more, but I also hear they've been tricky to support in general - I don't know to what degree we have something that works in the real world any more. (Although it is noted as experimental in the comment at least!)

Indeed: C++ modules are a mess. I'd like to keep whatever level of support we currently have, but it's pretty hard to figure out exactly what's broken across the various compiler versions - #61 should simplify that quite a bit...

[chrisnovakovic]

Force-pushing to rebase on master so we get the benefit of the new CI tests on this PR.

[chrisnovakovic]

I've given up trying to figure out whether Apple Clang supports C++ modules - the compiler's error messages say yes, but only with -std=c++20 or -fmodules-ts (neither of which compile //tests/modules); the preprocessor says no (neither __cpp_modules nor __cpp_lib_modules are set); some sources say they've been supported partially since 10.0.1, but give no indication of how that support can be enabled; some sources say yes, but only with -fmodules (an invalid option); other sources say no, not at all. It's incredible how badly-documented this all is.

I'm going to disable these tests by default unless they're re-enabled with a feature flag, and set that feature flag for the ubuntu_clang CI profiles.