[FEA] Add masking support to SequencePredictNext, SequencePredictLast, and SequencePredictRandom

[sararb]

🚀 Feature request

The current sequence transforms - SequencePredictNext, SequencePredictLast, and SequencePredictRandom - don't generate _keras_mask information for inputs and targets. This is because these transforms extract the targets from the input sequence and return new truncated inputs that don't contain the target information.

However, the Model output API uses mask information to adjust the shape of predictions and targets. To ensure that the Model output API can work with all supported sequence transforms, we should add masking support to all sequence transforms and rely on the Model output API to align the model's predictions and targets accordingly.

This would ensure that the model output API can handle a range of sequence transforms and produce correct outputs, regardless of the specific transform used.