Speed up Smagorinsky parameterisation

[Chris-Pedersen]

We have Smagorinsky implemented and working. However we are taking a large number of ffts:

        uh=-il*ph
        vh=ik*ph
        Sxx = torch.fft.irfftn(uh*ik,dim=(1,2))
        Syy = torch.fft.irfftn(vh*il,dim=(1,2))
        Sxy = 0.5 * torch.fft.irfftn(uh * il + vh * ik)
        nu = (self.constant * dx)**2 * torch.sqrt(2 * (Sxx**2 + Syy**2 + 2 * Sxy**2))
        nu_Sxxh = torch.fft.rfftn(nu * Sxx,dim=(1,2))
        nu_Sxyh = torch.fft.rfftn(nu * Sxy,dim=(1,2))
        nu_Syyh = torch.fft.rfftn(nu * Syy,dim=(1,2))
        du = 2 * (torch.fft.irfftn(nu_Sxxh * ik,dim=(1,2)) + torch.fft.irfftn(nu_Sxyh * il,dim=(1,2)))
        dv = 2 * (torch.fft.irfftn(nu_Sxyh * ik,dim=(1,2)) + torch.fft.irfftn(nu_Syyh * il,dim=(1,2)))
        ## Take curl to convert u, v forcing to potential vorticity forcing
        dq = -torch.fft.irfftn(il*torch.fft.rfftn(du,dim=(1,2)),dim=(1,2))+torch.fft.irfftn(ik*torch.fft.rfftn(dv,dim=(1,2)),dim=(1,2))

Given the relationship between the velocities and the streamfunction, and that we are mostly just taking derivatives here, there should be some speedups possible. And given that we need this Smag even for our "DNS", we should make this as efficient as possible.

[Chris-Pedersen]

So this is becoming more significant now due to #20