[feat] integrate dynamicemb table fusion (wheel 20260407.97b80bf)

[tiankongdeguiji]

Summary

• Bump dynamicemb wheel to the fused-storage build from NVIDIA recsys-examples PR #343 (commit 97b80bf). Multiple dynamicemb tables in the same TBE group are now backed by a single fused DynamicEmbStorage / cache / MultiTableKVCounter automatically.
• Adapt tzrec/utils/dynamicemb_util.py to the new API: the upstream _get_dynamicemb_options_per_table became a pass-through validator, so _to_sharding_plan now populates dim, per-shard max_capacity, embedding_dtype, and index_type directly. Remove the obsolete num_aligned_embedding_per_rank alignment shim and the matching monkey-patch — the field no longer exists on DynamicEmbTableOptions.
• Expose two new proto knobs on DynamicEmbedding:
  • bucket_capacity (default 128) — tunes hash table load factor vs. probe cost.
  • score_strategy = "NO_EVICTION" — keeps table capacity fixed once full, pairs with init_capacity_per_rank for growable tables.
• Doc updates for the wheel version and the new options; note that table fusion is automatic within a FeatureGroup.

Test plan

• python tzrec/tests/run.py for test_multi_tower_din_with_dynamicemb_train_eval — trains + evals a DIN model with dynamicemb features end-to-end on 2 GPUs.
• python -m unittest tzrec.tools.dynamicemb.create_dynamicemb_init_ckpt_test — verifies the init-ckpt workflow still produces a valid checkpoint that the trainer can consume.
• pre-commit run --files … — ruff, ruff-format, license header, mdformat all green.
• Smoke-tested build_dynamicemb_constraints with bucket_capacity=256 and score_strategy="NO_EVICTION" against the new wheel (yields a DynamicEmbTableOptions with the configured values).

🤖 Generated with Claude Code