Calling `pullback` twice in `test_rrule`

[mzgubic]

Our implementation of test_rrule calls pullback twice: once with the tangent to the primal output, and once with the thunk of the tangent.

Should we be doing this?

If the following conjecture is right:

The pullback is only ever called once when doing first-order AD. I.e. there is never a situation when a pullback is reused and called twice.

Then we are probably overly restrictive, as rules like #264 will not be able to be tested even if they are written with RuleConfigs that specify mutation must be supported.

But there might be a simple example to show the above conjecture is wrong, and then the status quo is fine.

[mzgubic]

From slack, by @oxinabox

Being able to do so is useful for computing Jacobean of functions with multiple outputs.