Fix NaN gradient in spectrogram

[pxl-th]

With power > 0 gradient will become NaN where spectrogram is ~0 (because of singularity at abs(0)).
Adding small epsilon fixes this issue (and the results match pytorch).

x = gpu(ones(Float32, 1024, 1, 1))
window = gpu(hann_window(1024)))
l, g = Zygote.withgradient(x) do x
    sum(spectrogram(x; n_fft=1024, hop_length=256, power=2.0, window)
end

PR Checklist

• Tests are added
• Documentation, if applicable

[mcabbott]

Sounds fine, a test would be great.