Add TabularPredictNext for causal item prediction task

[sararb]

Goals :soccer:

• Create the batch-block transformation operator TabularPredictNext for defining autoregressive targets by shifting the input sequence feature defined as target.

• Implement the BatchBlock TabularPredictNext.
• Implement TabulatSequenceTransform, the base class that checks the input sequences and target features are in the correct format.
• Implement TabularPredictNextModule responsible for shifting the targets and generating the related masking information.

Example of usage:

• For a group of sequences defined in sequence_schema, you can specify the target feature to use for creating the targets and all sequential inputs will be transformed accordingly.

  transform = TabularPredictNext(schema=sequence_schema, target="a")
  batch_output = transform(sequence_batch)

• You can also define multiple transforms for different sub-schemas:

      transform_1 = TabularPredictNext(schema=sequence_schema_1, target="a")
      transform_2 = TabularPredictNext(schema=sequence_schema_2, target="c")
      transform_block = mm.BatchBlock(
          mm.ParallelBlock({"transform_1": transform_1, "transform_2": transform_2})
      )
      batch_output = transform_block(sequence_batch)

[github-actions[bot]]

Documentation preview

https://nvidia-merlin.github.io/models/review/pr-1202