Add difsasrec

[TalCordova]

Summary

Implements DIF-SASRec (Decoupled Side Information Fusion for Sequential Recommendation)
from Xie et al., SIGIR 2022.

DIF-SASRec extends SASRec by decoupling item, position, and optional side-information
attention scores so each source independently shapes the attention distribution before
the softmax — rather than fusing them in the input embedding.

Key features:

• Three fusion strategies: sum, concat, gate
• Optional context towers: any FLOAT / FLOAT_SEQ item feature fields (e.g. LLM
  embeddings) can be added as extra decoupled attention sources
• Degrades gracefully to a pure DIF-SASRec (no side info) when context_fields: []

Files changed:

• recbole/model/sequential_recommender/difsasrec.py — model implementation
• recbole/model/sequential_recommender/__init__.py — registration
• recbole/properties/model/DIFSASRec.yaml — default hyperparameters
• tests/model/test_model_auto.py — integration tests (CE loss, BPR loss, concat fusion)
• tests/model/test_difsasrec.py — standalone unit tests (no full pipeline required)

Reference code: https://github.com/AIM-SE/DIF-SR

Test plan

• test_difsasrec — forward + CE loss, no side information
• test_difsasrec_with_BPR_loss — BPR loss path
• test_difsasrec_with_concat_fusion — concat fusion type
• All three fusion types (sum / concat / gate) × CE and BPR loss — via test_difsasrec.py
• Context towers (1 and 2 towers) with all fusion types
• Output shapes verified: forward, predict, full_sort_predict
• Context embedding weights confirmed frozen during training