Fixes and simplification in the Keras training engine.

- Explicitly disallow sample/class weighting in eager (it was never supported)
- Remove tests for it (which were actually ignoring sample/class weights)
- Make sample weight placeholders placeholder_with_default, and do not create all-ones numpy arrays to feed them when no sample weights are provided (this might lead to better performance)

PiperOrigin-RevId: 187121215