Pass `fused` arg to optimizer construction

[PatrickRMiles]

This will cause the grad unscaling and grad clipping to be fused with the optimizer step -- should speed things up.