Add a dont_dlopen flag to build_tarballs.

[maleadt]

This prevents BinaryBuilder from dlopen'ing library products during build, while still having the JLL doing so at run time. This matters for (surprise surprise) CUDA libraries, which commonly depend on libcuda, a library that will only be available at run time on a system with the CUDA driver installed. That means we don't want to load this library during the build on Yggdrasil, but we also don't want to set dont_dlopen on the LibraryProduct because we expect libcuda.so to be available while loading the JLL. If it isn't, that's detected by augment_platform!, which sets the platform tag so that no artifact will be loaded and no wrapper will be selected (and thus no library will be dlopened).

Ref JuliaPackaging/Yggdrasil#5701

[maleadt]

Coincidentally this is related to @giordano's recent work on not dlopening in case of ISA mismatches. Since we don't want to add CUDA-specific stuff to BinaryBuilder(Base), I guess the dont_dlopen keyword is still needed though.

[vchuravy]

Also helpful Flor libclang/LLvM and libclangex

Where we currently do this delightful nonsense https://github.com/Gnimuc/ClangCompiler.jl/blob/master/src/jllshim.jl

[maleadt]

Yeah I considered doing that, but JLLWrappers also does some LIB_PATH and dlopen_flag handling, and I didn't want to replicate all that.

[giordano]

Disentangling dont_dlopen in the generated JLL from the audit pass looks good to me, I'm however not happy about just adding extra random keyword arguments to build_tarballs(), but that's probably for another issue: #670