Racy bitcode compilation

I thought we already worked around this, using an atomic move-in-place: https://github.com/JuliaGPU/GPUCompiler.jl/blob/c15433ca68ee0fd52ae5241ad8d4c91df531801b/src/rtlib.jl#L145-L149

However, it looks like this still happens on CI, e.g. https://buildkite.com/julialang/metal-dot-jl/builds/947#018ee3aa-18e2-472c-b93f-f77451e3f87a

Start with the atomic move-in-place workaround in src/rtlib.jl at lines 145-149, then inspect the referenced Metal CI failure in Buildkite build 947. Determine why bitcode compilation can still race despite that workaround; done means the CI failure is addressed and compilation is no longer racy.

Written by the indexing model from the issue text.