This PR implements the fitted version of the superposition of atomic potentials guess (#564), which only requires three-center two-electron integrals.

Things that are still missing:

1. proper normalization of the contraction coefficients. The contraction coefficients should not be normalized by PySCF; instead, the input coefficients should be kept as is. They should sum to -Z, the negative of the nuclear charge. Then, before the integrals are calculated, the coefficients c_p just have to be multiplied with (\alpha_p / pi)^{3/2} so that the density primitives are normalized.
2. handling of ghost atoms; they should be skipped altogether in the auxiliary basis