Width-generic delta-swap pipeline shared by bitslice and inv_bitslice
across every Word impl. The same three-pass sequence inverts itself, so
bitslice and inv_bitslice invoke it identically.
Expand an 8-bit row pattern to a 16-bit row pattern by doubling each bit:
input bit i becomes output bits 2i and 2i+1. Branchless SWAR so LLVM
folds it to a single 16-bit immediate when b is a constant.