Pass keyword arguments to test_approx when checking thunks

[devmotion]

Just ran into a test issue where I need nans=true as keyword argument to isapprox (cf #220) but it was not forwarded when testing pullbacks with chunks.

[oxinabox]

This is definitely the correct behavior, regardless of if we do #220 or not.

I am not sure what is with the CI errors in Julia 1.0 on some plateforms but not others though.

[devmotion]

I am not sure what is with the CI errors in Julia 1.0 on some plateforms but not others though.

These test errors all seem to happen in checks that involve thunks. So I assume these tests were a bit brittle before but passed (usually) with the default atol=0 and rtol=sqrt(eps()) settings. With this PR, however, they are also run with atol = rtol = 1e-9 which probably uncovers these issues.

[devmotion]

A second observation is that the failing tests involve comparisons of Float64 and Float32. This might explain why it's an issue on some architectures only, and why atol = rtol = 1e-9 might be problematic.

Edit: And the test is stochastic on purpose it seems.

[devmotion]

Test errors should be fixed now. If there are no additional comments, I'll merge when the CR downstream tests pass.

[oxinabox]

And the test is stochastic on purpose it seems.

Right the stocasticity isn't on purpose, we just need to trick type inference into failing.
Which your solution does, while also removing stocasticity, so nice work 👍