ci: create release for Ubuntu 20.04

[alandefreitas]

This commit adds a new workflow that creates a Linux release on Ubuntu 20.04. This allows the release package to be used on more versions of Ubuntu. Since the C++ compilers we need are not available in Ubuntu 20.04, we use the clang executable from our own LLVM built from source to build the dependencies and MrDocs.

[alandefreitas]

While working on many things related to Antora + Mrdocs, I've been exploring this solution in parallel for about 10 days now. As each attempt might take a couple of hours, I've been tweaking CI here and there and seeing how it goes in parallel.

It's a good time to concede defeat. It's safe to say this is a bad idea to solve #547 because it's very unstable:

• It's overly complex / try-hard: To build MrDocs on Ubuntu 20.04 we need a custom-built compiler built with the tools Ubuntu 20.04 provides (clang + libcxx is the only option here) and then build LLVM again with this new custom toolchain. This involves many fine details that are hard to get right. It's also hard to maintain because we may need to adapt that workflow whenever MrDocs needs more new C++ features.
• It's slow: even with caching, building this custom toolchain and then building LLVM with it is much slower than the usual CI workflow. And every once in a while, the cache will need to be rebuilt, and building the toolchain twice might take over 2 hours.
• It's not general enough: This strategy allows us to support Ubuntu >=20.04 instead of Ubuntu >=22.04 or Ubuntu >=23.04. It doesn't generalize to other platforms, like the strategy CMake uses (APBuild, which is equivalent to a wrapper).
• It doesn't work properly. I've tried dozens of ways to make it work. If we don't statically link libc++ then we have the same problem generating the demos because we now have a compatible libc version but not a compatible libc++ on the machine using the release. But if we attempt to statically link libc++ we get other problems because clanglib doesn't support that properly.

Considering these limitations, the next idea I'll explore to solve #547 is a wrapper overriding system libs. This strategy is somewhat similar to what APBuild does but it's manually and more stable. It's similar to how CMake provides its binaries. This wrapper also seemed like a complex solution at first but it turns out to be much simpler and more effective than what we're trying to do here. I'll describe the proposed solution in #547.