well I just derived them empirically-- you can test the coefficients yourself by plugging the values into filtergraph and see what sort of a curve you get. I did test the filter against a class-A spl meter to see how accurate it was and it seemed to be within less than 1 dB margin of error, which was adequate for my purposes at the time.
as for the other question, you could design a set of cascading fft modules using pfft~, each of which only samples bins 4 to 6 of a different window-size (ie 4096, 2048 ......128, 64), one for each octave
that code is about equal loudness, not A-weight which are completely different things, and it's also written in matlab, which I don't know...af probably refers to amplitude (a) at corresponding frequency (f), but that's just a guess
btw my previous post should read 'bins 3 to 5", not '4 to 6", if you want pseudo 1/3 octave analysis