Get fully non-allocating for zero backends

[gdalle]

At the moment there are still broken tests.
Some of them seem to depend on the benchmarking functions themselves, like multiderivative! with AutoZeroReverse.
When you remove the function evaluation, it passes, even though the function in every mutating scenario is supposed to be allocation-free.
Even weirder, when you remove just the second line of f!_scalar_vector, it passes as well.
Maybe something to do with inlining?

[gdalle]

So I figured it out in #75

I think small enough functions were not affected by this because they were inlined