make sparsemax nan and infinity safe

logits that are -inf will be given 0 probability and logits that are
inf will result in a nan output. Likewise if all logits are -inf the
output will also be nan.

This is done by using where operators, mostly because 0 * inf = nan
and x/0 = sign(x) inf following the IEEE 754 standard. However these
results are not mathematically correct in the context of the sparsemax
algorithm.

Fixes: https://github.com/tensorflow/tensorflow/issues/15564