More accurate softlif in tensorflow

[hunse]

This is a change that I also made in nengo_extras (see here).

Basically, for small j, z = log(1 + exp(j/sigma))*sigma = sigma*exp(j/sigma) since log(1 + x) = x for small x. Then, log(1 + 1/z) = log((z + 1) / z) = log(1 / z) since z is also small and z + 1 = 1. Finally, log(1/z) = -log(z) = -log(sigma*exp(j/sigma)) = -j/sigma - log(sigma).

[drasmuss]

Pushed some minor fixups to the comments, as well as some updates I made to benchmarks.py while I was testing this. For reference, this seems to result in about a 5-10% performance hit in training/inference on the benchmarks, but that seems like a reasonable price to pay for the increased accuracy.