Fix NC metadata, training flow, encrypted aggregation, and smoke test compatibility

[Pulserius]

Summary

This PR fixes several NC workflow and test issues:

• Makes NC metadata handling more robust for Hugging Face-loaded trainer data:
  • preserves/saves global_node_num and class_num
  • handles empty train/test label tensors safely
  • avoids recomputing class_num from partial trainer labels when authoritative metadata is available
• Moves NC federated round execution into Server.train() and has run_NC() reuse that path.
• Fixes encrypted parameter aggregation metadata handling across model layers.
• Removes unused server-side encryption helpers and unused get_local_feature_sum_og().
• Restricts NC num_hops to the currently supported modes: 0 and 2.
• Masks encrypted feature sums per trainer before sending them back.
• Fixes duplicate trainer test metric appends.
• Updates smoke/OpenFHE tests for the merged main branch compatibility.

Validation

Ran focused regression tests for:

• NC metadata fixes
• NC server training flow and encrypted aggregation
• NC hop validation, encrypted feature masking, and unused helper removal
• duplicate test metric append fix
• merged-main smoke tests and OpenFHE smoke tests on EC2 Linux

Key test groups passed:

python -m pytest -q tests/unit/test_federated_methods.py::TestResolveNCClassNum tests/unit/test_federated_methods.py::TestResolveNCGlobalNodeNum

python -m pytest -q tests/unit/test_trainer_class.py::TestLoadTrainerDataFromHuggingFace tests/unit/test_trainer_class.py::TestTrainerGeneral::test_get_info tests/unit/test_trainer_class.py::TestTrainerGeneral::test_get_info_handles_empty_labels

python -m pytest -q tests/unit/test_utils_nc.py::TestSaveTrainerDataToHuggingFace

python -m pytest -q tests/test_smoke_unit.py

python -m pytest -q tests/test_smoke_e2e.py::test_smoke_plaintext_default_norm_none tests/test_smoke_e2e.py::test_smoke_plaintext_norm_sym_opt_in

python -m pytest -q test_openfhe_smoke.py tests/test_threshold_ckks_min.py

[yh-yao]

looks good