Optimize opt_expr evaluation ordering

[KrystalDelusion]

Feature Description

While working on #4899 I discovered that opt_expr has optimizations for folding cells with constant outputs (e.g. when a shift cell has two constant inputs), but these don't get checked until after the optimizations for cells with one constant input. This means for example, that if a shift cell has two constant inputs, it will first replace the cell with a wire connecting the relevant bits of A directly to Y, which then requires a separate operation to optimize the new wire. It seems to me that reordering the opt_expr::replace_const_cells() to perform the FOLD_NARG_CELL() checks first could be an improvement to performance (albeit a slight one).