Add support for `Nx.rfft` and `Nx.irfft`

[msluszniak]

I would be a great addition. Currently, a bottleneck for me is such kind of code:

xfft = Nx.fft(xw, length: nfft, axis: 1)[[.., 0..cutoff]] |> Nx.new_axis(-1)
outf = xfft * filter[[.., 0..cutoff, ..]]
flipped = Nx.conjugate(outf[[.., 1..-2//1, ..]]) |> Nx.reverse(axes: [1])
outw = Nx.concatenate([outf, flipped], axis: 1)

If these operation will be implemented efficiently, the code above can get me a huge speed-up:

xfft = Nx.rfft(xw, length: nfft, axis: 1) |> Nx.new_axis(-1)
outf = xfft * filter[[.., 0..cutoff, ..]]
temp = Nx.irfft(outf, length: nfft, axis: 1)

[polvalente]

I think it's ok if we could implement this. Do you want to take this on? Let's wait for #1709 to get merged and we can do it!

[msluszniak]

I think it might be hard for me to implements this because of lack of time. If I would have time before merging #1709 I'll let you know that I'm implementing it, so we don't overlap :)

[polvalente]

rfft/irfft will directly overlap with that PR, so we'll have to merge them afterwards

Don't worry about rushing this, though. We can tackle it after the PR