Streamline the autoformat workflow to enable version pinning for all platforms

[rdw-software]

TODO:

• Cleanup and rebase
• Simplify the workflow
• Eliminate need to define env variable for clang-format
• Document the process for installing and running formatters (Tracked separately)

---

The autoformat workflow currently doesn't pin the version of clang-format which is supposed to be used. That's problematic because both MSYS2 and Homebrew install the "latest" version, but their definition of "latest" can differ quite a bit.

With this patch, the version is pinned for all three platforms and can manually be updated to track MSYS. It's still possibe for the versions to diverge, when MSYS deploys a new major release and the workflow hasn't been updated. In that case, there might be breaking changes which should lead the workflow to fail (can update version to re-sync). Otherwise it "just works".

As a side effect, the autoformat script will now fail unless the clang-format version is set via EVO_AUTOFORMAT_VERSION. This is slightly annoying because it makes running the script more cumbersome, but it can be exported automatically. If that's too inconvenient I'll have to think of a better solution later. (Stored in autoformat.env now)

I don't particularly like the increase in complexity, but formatters randomly messing up the code isn't helpful either.

[rdw-software]

Not happy with how much unneeded complexity is introduced by having three different sources for clang-format. The binary name is also quite inconsistent and renaming it as part of the workflow might work, but doesn't help with local development. Ideally, would be able to download just clang-format from GitHub releases. Alas, it is part of the bloated 1.5 GB tarball.

Stylua isn't a problem, thankfully, as their releases are much less chonky and can easily be fetched on the fly. The action doesn't work in MSYS (GitHub limitation?), which is slightly annoying and could be revisited. For now, focus on the LLVM setup...

A few potential ideas:

• Use alternative sources like chocolatey for Windows (still depends on third parties, may not have the right version)
• Build clang-format from source and use the latest HEAD/some specified version (still takes ages and wastes bandwidth)
• Fetch LLVM releases from GitHub, create a release for clang-format via GitHub Actions (in another repo), fetch that
• Just give up and again have one single source of truth - the official LLVM APT repo might have the fastest install time?
• Fetch it once and store in GitHub's cache, then use that for a given version (can save version in env file and hash it)
• Accept having multiple versions and create alternative formatter configs for each (doesn't seem like a good idea)
• Build or fetch manually a compatible version, host it externally/via artifacts, and use it (don't like the security headaches)

I think the PPA is the best option for Linux and it does provide recent versions. Homebrew/MSYS are quite slow/large. Can't use the Linux releases anyway due to GLIBC issues; so there may not be a good way to get consistency and speed here?

[rdw-software]

After much iteration, this is the simplest solution I could come up with that fulfills all the requirements:

• A single clang-format version is used for all platforms
• The version can be pinned in one central location
• Whichever version is currently used will be tracked by git
• Doesn't consume resources excessively thanks to caching
• No reliance on third parties, other than the LLVM organization
• Uses the same script for both CI runs and local development
• No excessive branching/other accidental complexity
• Bonus: Approach can also be used to install clang-tidy (later)

Downloading a huge release tarball just for the tiny formatter feels like overkill, but it's reasonably fast in the CI.