When using isapprox to compare arrays, the current definition, introduced in #12472, seems to have some (IMHO) unideal behavior in the presence of NaNs:

julia> [0.01311489462160816,Inf] ≈ [0.013114894621608135,Inf]
false

julia> all(map(isapprox, [0.01311489462160816,Inf], [0.013114894621608135,Inf]))
true

This is how isapprox is currently defined for arrays:

function isapprox{T<:Number,S<:Number}(x::AbstractArray{T}, y::AbstractArray{S}; rtol::Real=Base.rtoldefault(T,S), atol::Real=0, norm::Function=vecnorm)
    d = norm(x - y)
    return isfinite(d) ? d <= atol + rtol*max(norm(x), norm(y)) : x == y
end

Whenever norm evaluates to NaN, which happens for example when x and y both have Inf in a corresponding index, the comparison falls back to checking exact equality using ==. That will be false if there are finite, approximately equal floating point values in other corresponding indices in x and y. This is causing a test failure in HypothesisTests.jl.

@stevengj was against checking approximate equality element-wise for arrays but I'm not entirely clear on why. Is there anything that can or should be changed in the current implementation, or is there some other way I can get around the test failure?