This is because of the slightly different way cftool has defined the gaussian equation for the fit, and it ends up multipling the c1 coefficient by a factor of sqrt(2) from the true value of the standard deviation.
The equation for FWHM is
FWHM = 2*sqrt(2*log(2))*sigma
For fitting, MATLAB uses
f(x) = a1*exp(-((x-b1)/c1)^2)
And wherever you're reading the FWHM equation defined the gaussian as
f(x) = (1/sigma/sqrt(2*pi))*exp(-(x-mu)^2/(2*sigma^2))
By equating the two, you'll find that
Therefore the FWHM equation becomes
FWHM = 2*sqrt(2*log(2)) * (c1/sqrt(2))
FWHM = 2*sqrt(log(2)) * c1
And in fact, when you use this new equation, you get FWHM = 21.78, which is accurate.