[llvm] Don't use libcalls for i128<->fp conversion on 32-bit targets (PR #196330)

[Nikita Popov via llvm-commits]

nikic wrote:

libgcc has __mulbitint3 and __divmodbitint4 for `_BitInt`, but it does not define builtins for conversion functions between `_BitInt` and FP. (Incidentally compiler-rt does not implement these builtins and we currently emit inline expansions for mul/div instead.)

https://github.com/llvm/llvm-project/pull/196330