No mention of Padé approximants? I'd just use a Chebyshev approximation transformed into a Padé approximation, then perturb the coefficients to optimize maximum error.