You can’t, because that’s not how a multidimensional FFT of real inputs works. You can only save ≈half the data and computation regardless of the dimensionality. That is, you can only cut one dimension in half, and FFTW for convenience chose the first dimension (in Julia column-major order, corresponding to the last dimension on C).
See this explanation from the FFTW manual:
Conceptually it is transforming along each dimension, and then discarding a redundant half of the output.
In practice, it first performs the rfft along the first dimension, which produces complex data of half the size, and then transforms along the remaining dimensions — but since the result of the first transform was complex, the remaining computations are ordinary complex FFTs.