isapprox is incorrect for tolerances beyond standard precision

The following is incorrect:

```julia
julia> T = Double64
Double64 (alias for DoubleFloat{Float64})

julia> x = Double64(0.9, 1e-20)
9.00000000000000022214460492503130852e-01

julia> abs(x - T(9)/10)
2.22144604925031320405193304841894479e-17

julia> isapprox(x, T(9)/10; rtol=1e-22, atol=1e-22)
true
```

If I understand the internals, the isapprox function uses `HI(x)===HI(y)` as a shortcut for the test, which doesn't look right.