divide-by-zero in torch.ifft()

in `utils.activations_to_audio()`, there's a stanza for scaling the signal to the open range [-1, 1]. This stanza assumes that `signal.max() - signal.min() > 0`, and so can throw a divide-by-zero error.

Start in utils.activations_to_audio() and inspect the signal-scaling stanza around the signal.max() - signal.min() calculation. Reproduce the constant-signal case, determine the expected output from the surrounding audio conversion behavior, and add or update a focused test so the function no longer raises a divide-by-zero error.

Written by the indexing model from the issue text.