delay scaling of ffts until after the transforms.

As seen here, the results can be order of magnitudes off if the scaling happens upfront.
https://gist.github.com/pavanky/dc64fc83e0fa298e942a3e0ca07ccd15

Start with the linked gist and reproduce the FFT case where upfront scaling produces results orders of magnitude off. Trace the Python FFT entry point and its transform/scaling path; done means scaling is applied after the transforms and the reproduced results match the expected values.

索引モデルが issue の本文から書いたものです。