glsl/lower_instructions: fix regression in dldexp_to_arith