Lint SIMD code, fix generic expectation failures

[clarfonthey]

Closes #691.

This code ends up using a cursed bash feature (associative arrays) to help people running the lints on their own machines. Since we don't actually need a host compiler to lint the code, we can just download the necessary targets with rustc, with an exception carved out for whichever SIMD implementation the host target supports, to avoid downloading unnecessary targets (e.g. aarch64-unknown-linux-gnu when you're running on a Mac, for example).

Side note: this could be done with POSIX sh by replacing the associative arrays with variable expansion (creating new variables instead), but I couldn't be bothered when pretty much everyone has a copy of bash available.

[clarfonthey]

cc @xtqqczze since you were the one who brought this up initially, and I assume you can run the script to verify that it correctly identifies that your target uses NEON.

[xtqqczze]

I run the script, it appeared to run successfully, and I saw "Host target supports NEON" in the output.

FYI: If you wanted, you could specify something like the following in the GitHub Actions file:

      shell: bash
      run: ./ci/tools.sh

[clarfonthey]

Eh, I figured I wouldn't bother messing with the CI config too much. The scripts aren't executable which is annoying, but, whatever.

Might change my mind and fix it later, though. :p

[xtqqczze]

I just noticed the only target for miri test in CI is x86_64-unknown-linux-gnu...

[xtqqczze]

We should probably include s390x-unknown-linux-gnu in some check, so as to have at least one big-endian target.

[clarfonthey]

Yeah, not sure what the best solution for the miri tests would be, since it depends a lot on CI capacity. We already have quite a lot of jobs running and it's not exactly clear what the project is currently able to support. Ideally, we'd have miri run every test suite that's run normally, but that would be quite a lot of overhead.