default test tolerances assume Float64 precision and numbers ~ 1

[stevengj]

The default test tolerances here seem wrong:

ChainRulesTestUtils.jl/src/testers.jl

Line 17 in 6925da1

function test_scalar(f, z; rtol=1e-9, atol=1e-9, fdm=_fdm, fkwargs=NamedTuple(), check_inferred=true, kwargs...)

1. The default rtol = 1e-9 assumes double precision. Better to follow isapprox here and use sqrt(eps(float(T)))?
2. By setting a nonzero default atol you are implicitly assuming quantities of order unity. This corresponds to an absurdly large tolerance if the functions happen to be scaled to have very small magnitude. Better to follow isapprox here and default to atol = 0.

[oxinabox]

This makes sense.
It should be a negligible change, and we can use the type of the input to determine the type we should use for rtol, at least for the test_scalar case. Might be a little more fiddly for the test_rule and test_frule.

We should check if this breaks anything downstream, but assuming it doesn't happy enough to see this in a bug fix release.
Otherwise we might have to do more thinking