[doc] How to upscale FFT input and avoid overflow?

[puzrin]

Typical use case is to apply fft_fft() to ADC samples (8-12 bits). User should upscale input as much as possible, but avoid saturation/overflow.

CMSIS DSP uses Q31 input and downscale result automatically. As far as i understand, SYLT-FFT don't downscale result (to improve performance), and user should select proper input scale manually.

Is there a clear formula/table, how to select maximal possible input scale for concrete fft sizes (32-4096)?