Reconsider Loss generic parameter

I brought this up with @JimClarke5 but wanted to get some wider comments on it.

The generic parameter for Losses doesn't seem right:

<T extends TNumber> Operand<T> call(Ops tf, Operand<? extends TNumber> labels, Operand<T> predictions, Operand<T> sampleWeights);

It binds the loss value, the predictions, and the weights to the same data type. There is no relation there: it's easy enough to imagine float-weighted cosine similarity of one hot vector predictions, which would return a float. The only necessary data type parameter I can think of is the return type, and that should be on the class.

Start with the Loss generic method signature shown in the issue and inspect the surrounding Loss API and its callers. Review the existing comment discussion for agreement on the type relationships. Done means the generic design is decided and the affected API and callers consistently reflect that decision.