Bunch of GPU detection work

[Artoria2e5]

• Add myfnmatch.c/h for simple wildcard matching. Unix systems have fnmatch() but Windows doesn't, so we provide a version from musl.
  • If I was any greedier I would import gnulib and use regex from there. But that would be a lot more lines!
• Use fnmatch() to:
  • Match GPU models, which is what you get on not-Linux. Should improve detection on Windows and macOS.
  • Match gfx-targets, which is what you get on Linux. Should cover more AMD GPUs. Also added matching for CDNA, which are chonkier versions of GCN5 (1:1 FP64 for all CDNA, 1:1 INT32 MUL since CDNA2)
• Cover RDNA4 and RDNA3.5 as using RDNA3 features. Should help with mfakto 0.16-beta.3 and no support for RDNA4 ? #43 but definitely test before closing.
• Edit the ini comments to match how we actually use the architectures. Especially APU, since no new APU is put in that bucket.

Compiles on mingw-w64 (MSYS2 "Mingw64") with the following caveat:

• No LTO enabled. Using LTO causes a symbol resolution error between C++ mangled and C non-mangled symbols on fnmatch when linking. This is most likely a toolchain bug, since fnmatch is built under C and referenced with extern C.

[ixfd64]

Your code was only compiling on Windows. I've fixed the macOS and Linux build errors.

[ixfd64]

I don't see any issues with your changes. However, I'd like others to take a look at it as well if possible.

In the meantime, I've improved the code formatting and INI file comments. Hope you don't mind. :-)

[Artoria2e5]

These are all good changes! I am thinking about removing the mention of Rtl... whatever the glob thing on Windows is though, because with the use of brackets in the patterns it just won't work. I initially wanted to use simpler patterns (or at last some sort of two-stage matching where the complex character-sets are just replaced by ?), but that turned out unnecessary with how small a proper fnmatch is.

[ixfd64]

I think the ideal solution would be to install external dependencies via a package manager. However, that's not really an option in this case because musl doesn't seem to be in the vcpkg repo.

[N-Storm]

Wouldn't PathMatchSpecExA() work as a replacement for fnmatch() on Windows?

[Artoria2e5]

The better replacement is the Rtl stuff in ntdll. The path one has, well, path-specific idiosyncrasies especially with 8.3 that we don't need. But even "better" isn't good enough as I've decided to be greedy and reach for the [] part of POSIX glob patterns.

[N-Storm]

I don’t see any obvious issues. I didn’t check the complex code in myfnmatch.c in detail, but it matches the upstream version from musl, taking into account the declared changes that remove unnecessary functionality for some flags.

What does concern me, though, is that the build with LTO on MinGW doesn’t work. It’s probably worth examining this further, but I don’t have enough spare time right now to look into it. So I’ll approve the PR for now, but this should be revisited later.

[Artoria2e5]

WAIT A SECOND fnmatch matches the whole string, so if anything is to be matched a bunch of * is needed woah great job me now I have a lot to explain