Don't override precision directly in the QKeras optimizer

Created by: vloncar

output_rounding_saturation_mode optimizer directly manipulates precision object of the output variable of the node, which can trip up the later optimizers and make them skip the type conversion. It was observed in testing of accelerator backend with fifo depth optimization (by @thesps) as well as some GNN models (by @sznajder) This change ensures the update of precision also updates the entire type. Eventually we will make the precision a read-only property to ensure this type of error doesn't affect future optimizers, but this kind of change requires a bit more work.