This shows up to 5% less instruction counts on multiple benchmarks, and up to
19% wins on the -j1 wall times for rustc self-compilation.

We can afford to spend the extra cycles building LLVM essentially once more for
the x86_64-unknown-linux-gnu CI build today. The builder finishes in around 50
minutes on average, and this adds just 10 more minutes. Given the sizeable
improvements in compiler performance, this is definitely worth it.

This skips bumping Windows sccache because we run into compilation failures when
doing so (-m32 not supported by clang-cl?). Not clear on cause, but seems
easiest to just hold back.

This should avoid PGO-related failures encountered on Linux, and more broadly
seems like a good idea on other platforms as well (though it is likely not
necessary right this moment).

This will ensure the optimized LLVM is used for local builds after rebasing
immediately, rather than needing to wait for a future LLVM bump.