radeonsi: add a debug flag for unsafe math LLVM optimizations