LLVM: account for register type in funPrologue

We were not properly accounting for the live register type of global registers in GHC.CmmToLlvm.CodeGen.funPrologue. This meant that we could allocated a register at type <4 x i32> but try to write to it at type <8 x i16>, which LLVM doesn't much like.

This patch fixes that by inserting intermediate casts when necessary.