On the accuracy and performance of the lattice Boltzmann method with 64-bit, 32-bit and novel 16-bit number formats