Use active MPU for AutoEP sequence-parallel size

[tohtana]

Summary

• Fixes the Codex bot review comment: Add AutoEP deepspeedai/DeepSpeed#7938 (comment)
• Reads AutoEP sequence-parallel size from the active engine MPU when it exposes get_sequence_parallel_world_size, before groups.mpu is populated.
• Keeps the existing groups._get_sequence_parallel_world_size() fallback for non-MPU paths.
• Adds unit coverage for MPU-first AutoEP validation/group creation and the groups-helper fallback.

Intended Target

This PR is intended to merge into tohtana/add_autoep, the branch behind upstream deepspeedai/DeepSpeed PR deepspeedai#7938.

Tests

Environment: /mnt/local_storage/autoep_transformers_matrix_20260513_fixed/venv-5.8.1
Transformers: 5.8.1

• timeout 300 /mnt/local_storage/autoep_transformers_matrix_20260513_fixed/venv-5.8.1/bin/python -m pytest -q tests/unit/moe/test_autoep_unit.py::TestAutoEPConfig::test_configure_expert_parallel_uses_engine_mpu_sequence_parallel_size tests/unit/moe/test_autoep_unit.py::TestAutoEPConfig::test_autoep_sequence_parallel_size_falls_back_to_groups_helper
• timeout 900 /mnt/local_storage/autoep_transformers_matrix_20260513_fixed/venv-5.8.1/bin/python -m pytest -q tests/unit/moe/test_autoep_unit.py

Both passed.