I have always mean centered data when concerned with autocorrelation
with standard OLS regression.
I have not seen anything with regard to mean centering with FGLS but
should I do so before running data?
SInce altering the means of Y and X by a constant only affects the
intercept in regression (and centering sets it to zero) it has
nothing whatsoever to do with the serial correlation of the
disturbance process. Centering the data merely causes the regression
surface to pass through (0,0 [0,0,0,...]). It will have no effect on,
say, a consistent estimate of \rho in an AR(1) error scheme.