sd_fft: Add more documentation

[user202729]

Mostly trying to understand the algorithm myself.

The notation

$A$ independent contiguous $B$-layer transforms with length $(C, C/2, \dots, C/2^{B-1})$ and mask $(D‥E)$

is probably not standard, but I think it's understandable enough. I didn't formally define it either, but (I hope…) the formal definition is clear from the examples.

(It seems to me that mathematicians prefer formal definitions, implementors prefer "definitions" by examples. We're both. Upside of formal definition is it can be understood eventually given enough effort, downside is if the examples can be understood then it would take more effort to read the formal definitions.)

While writing the documentation, I notice that the internal variants do not make sense with S != 1, so I simplify it by deleting that argument. Should translate to a tiny performance improvement.

The numbering of the layers might have been backward (it's internally consistent, problem is it may be bad notation).