Make the isapprox implementation closer to the LinearAlgebra implementation

[eliascarv]

The Unitful implementation of isapprox(::AbstractArray, ::AbstractArray) function has some differences compared to the LinearAlgebra implementation.
Link: https://github.com/JuliaLang/julia/blob/master/stdlib/LinearAlgebra/src/generic.jl#L1879-L1891

For example, the comparison in Unitful is:

d <= atol + rtol*max(norm(x), norm(y))

But in LinearAlgebra is:

iszero(rtol) ? d <= atol : d <= max(atol, rtol*max(norm(x), norm(y)))

Another difference is in the use of the Base.rtoldefault function.
The Base.rtoldefault(x, y, atol) function returns the default rtol only when atol is less than or equal to zero, otherwise it returns zero.
But, in Unitful, the last argument of Base.rtoldefault is always zero:

rtol::Real=Base.rtoldefault(T1,T2,0),
atol=zero(Quantity{real(T1),D,U1})

That is, the rtol value is always be different than zero, even when atol > 0.
Below is the LinearAlgebra implementation for comparison:

atol::Real=0,
rtol::Real=Base.rtoldefault(promote_leaf_eltypes(x),promote_leaf_eltypes(y),atol)

The last difference is that the Unitful doesn't forward the nans keyword argument.

These different implementations result in different behaviours:
master branch:

julia> x = [-48.044483f0, -18.32653f0]
2-element Vector{Float32}:
 -48.044483
 -18.32653

julia> y = [-48.04448f0, -18.326504f0]
2-element Vector{Float32}:
 -48.04448
 -18.326504

julia> isapprox(x, y, atol=1f-5)
false

julia> isapprox(x * u"m", y * u"m", atol=1f-5u"m")
true

This PR adjusts the implementation to be closer to the LinearAlgebra implementation:

julia> x = [-48.044483f0, -18.32653f0]
2-element Vector{Float32}:
 -48.044483
 -18.32653

julia> y = [-48.04448f0, -18.326504f0]
2-element Vector{Float32}:
 -48.04448
 -18.326504

julia> isapprox(x, y, atol=1f-5)
false

julia> isapprox(x * u"m", y * u"m", atol=1f-5u"m")
false

[eliascarv]

@sostock, PR ready for final review and merge.
And, after the merge, can you release a new version please?

[sostock]

There should be some tests for the new behavior. Suggestions:

@test isapprox([1.0*m, NaN*m], [nextfloat(1.0)*m, NaN*m], nans=true)
@test !isapprox([1.0*m, NaN*m], [nextfloat(1.0)*m, NaN*m], nans=false)
@test !isapprox([1.0m, 2.0m], [1.1m, 2.2m], rtol=0.05, atol=0.2m)
@test !isapprox([1.0m], [nextfloat(1.0)*m], atol = eps(0.1)*m)

If I didn’t make a mistake, all of these tests should error/fail on the old version and pass with this PR.

The scalar isapprox(::AbstractQuantity{T,D,U}, ::AbstractQuantity{T,D,U}) where {T,D,U} method should be changed to use ustrip(absoluteunit(unit(y)), atol), (and tests should be added for that as well), but that is unrelated to the changes in this PR and can be done later.

After this PR is merged, I will make a new release (a minor release, since in my opinion the addition of the nans keyword qualifies as a feature).

[eliascarv]

@sostock ready for merge