fix wrong partials multiplied in FMA

[KristofferC]

The methods gave the wrong partials but this was not noticed in the tests because == only compares the value part:

julia> x, y, z = Dual(1,1,2,3), Dual(2,4,5,6), Dual(3,7,8,9);

julia> x*y + z - fma(x, y, z)
Dual(0,-3,-3,-3)

Fix the bug and introduce a dualapprox function to test approximate results of both the value and partial part:

julia> x, y, z = Dual(1,1,2,3), Dual(2,4,5,6), Dual(3,7,8,9);

julia> x*y + z - fma(x, y, z)
Dual(0,0,0,0)

cc @tpapp

[mlubin]

I'm guessing the commit of partials.jl.mem was not intentional?

[KristofferC]

Jeez, it's up for 15 seconds and someone notices.. :P

[jrevels]

The (outdated-ly named) test_approx_diffnums defined below is what is used for this in the rest of the tests. Can you reuse/rename/rewrite that function instead?

[KristofferC]

Sure, didn't notice it.