That’s somewhat of an ugly parametrization; using
x = \frac{a}{a+b}
and
y = \sqrt{a+b}
you can sample x and 1/y uniformly, and use the transformation
a = xy^2
b = y^2 - xy^2 = (1-x)y^2
to get an equivalent prior from uniform samples. See Hierarchical AB testing in PyMC3 - #2 by chartl