Avoid division by zero in TPR (and similar metrics)

[tmke8]

TPR is defined like this:

return t_pos / (t_pos + f_neg)

If both t_pos and f_neg are 0, then numpy produces NaN. But I'd argue that the result should be 0 instead.

cc @olliethomas @MylesBartlett

[olliethomas]

I know we've spoken about this before.

Conversation from 24th November 2020

oliver 15:40
maybe 0/0 should be 0 :man-shrugging:
thomas_k 15:40
I think it’s more informative to report this as nan

and I know we'd discussed it before then too.... Maybe. we should just make it 0

[tmke8]

hm, I see... maybe my past self was right.

[tmke8]

I think we should still avoid the division by zero because it produces annoying warnings. We can still return NaN though.