NaNs when using ForwardDiff with RodriguesVec

For example:

julia> using ForwardDiff, Rotations

julia> ForwardDiff.derivative(0) do x
           rotation_angle(RodriguesVec(x, 0, 0))
       end
NaN

This comes out of the way the rotation angle is computed, which involves sqrt(rv.sx * rv.sx + ...), and which fails because the derivative of sqrt at 0 is inf (from above).

However, there is a correct answer for this derivative (it's 0), so I wonder if we can make the computation a bit more robust. For example, the generic LinearAlgebra.norm does the right thing:

julia> ForwardDiff.derivative(0.0) do x
           rv = RodriguesVec(x, 0, 0)
           norm([rv.sx, rv.sy, rv.sz])
       end
0.0

Reproduce the ForwardDiff.derivative example using rotation_angle(RodriguesVec(x, 0, 0)) and inspect the rotation_angle computation that takes the square root of the RodriguesVec components. Compare its behavior with LinearAlgebra.norm at zero. Done means the derivative at zero returns 0 rather than NaN while preserving the expected rotation angle.