Document choosing different loss functions

[stsievert]

As discussed in #218, it's hard to choose the loss automatically (categorical cross entropy loss is best for independent labels, binary cross entropy is okay for multiple labels).

Motivation for filing this issue: #218 (comment)