By switching to Scalars that are backed by explicitly-sized APInts we can avoid a bug that increases the buffer reserved for a small piece to the next-largest host integer type.
This manifests as "DW_OP_piece for offset foo but top of stack is of size bar".