Please use multithreaded xz

[froody]

My team are constantly complaining that llvm takes over a minute to extract, even if the download is cached. This is due to bazel using single-threaded java to extract .xz archives. It's possible to use an external xz toolchain which uses multithreading, see google-ml-infra/rules_ml_toolchain@50a9847 for examples

[dzbarsky]

It's even better to recompress the archives with zstd, you can get smaller archives that are even faster to extract.