Clean up model input names for consistency

[mattdangerw]

This proposes a few changes to the naming our our model and layer inputs
and outputs.

1. Rename input_ids -> token_ids for bert/roberta.
   Everything is an "input", including the segment id input, so I don't
   think input is a helpful naming prefix in this case.
2. Rename input_mask -> padding_mask for bert/roberta.
   This matches the name of the variable for the transformer
   encoder/decoder argument.
3. Rename tokens -> token_ids for MLMMaskGenerator.
   This layer only operates in id space, so I think token_ids is more
   descriptive and consistent with above.

[jbischof]

Looks good in general but I'm confused by the padding_mask label. Aren't we also masking the holdout words in pretraining and not just the padding?

Does the examples/bert/ code still work? I see it still uses input_ids here.

Finally, there's one other naming inconsistency you could mop up in sine_position_encoding.

[jbischof]

LGTM! Added some comments for consideration but trust your judgment.

[mattdangerw]

@jbischof Thanks! And good call I will rename in examples/bert to match.

Re padding_mask and the MLM masked words, that's actually a different mask I think. There's actually a [PAD] token and a [MASK] token in the bert vocab. During pretraining, [PAD] will be used only to pad sequences, and [MASK] will be used for the MLM labels.

So I do think it is accurate to use padding_mask as a name here. Other options are mask or token_mask. I am actually fine with any of those. I was just choosing padding_mask because that the name we chose for our TransformerEncoder/TransformerDecoder https://keras.io/api/keras_nlp/layers/transformer_encoder/#call-method.

[fchollet]

LGTM