[bug] Dont always sample Float64's

[lrnv]

          I think this is an upstream problem - the `copula` (and hence also `input_distribution`) lost the type information and generate samples of type `Float64`.

Originally posted by @devmotion in SciML/GlobalSensitivity.jl#158 (comment)