Add BERT4Rec to KerasHub

[divyashreepathihalli]

No description provided.

[harshaljanjani]

I wish to take up this issue.

Key Differences Between BERTBackbone and BERT4RecBackbone:

Embeddings:
BERTBackbone: Uses token, positional, and segment embeddings.
BERT4RecBackbone: Uses only token and positional embeddings (no segment embeddings).
Source: BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer (Section 3.4: Embedding Layer).

Pooling:
BERTBackbone: Outputs a pooled representation (via [CLS]) for classification.
BERT4RecBackbone: Outputs the entire sequence representation for masked item prediction (pooling was definitely under consideration but doesn't seem to be enforced in the final rendition).
Source: https://github.com/FeiSun/BERT4Rec/modeling.py#L220

If there are any others, please let me know, but to my knowledge these changes are sufficient to make it viable for BERT to support sequential recommendation, as described in the paper.