Remove AutoEP backward loss multiplier

[tohtana]

Summary

• Removes the AutoEP-only ep_size loss multiplier from DeepSpeedEngine.scale() and DeepSpeedEngine.backward().
• Removes the now-unused AutoEP backward scaling state/helper.
• Updates the AutoEP ZeRO-2 gradient parity test so all-gathered expert gradients are no longer divided by ep_size.

Context

Addresses Codex bot review comment: deepspeedai#7938 (comment)

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

Testing

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

Transformers version:

5.8.1

Commands run:

/mnt/local_storage/autoep_transformers_matrix_20260513_fixed/venv-5.8.1/bin/python -m compileall -q deepspeed/runtime/engine.py tests/unit/moe/test_autoep_grad_parity.py
flock /tmp/ds_gpu_test.lock -c 'PYTEST_XDIST_WORKER=gw1 DS_DISABLE_REUSE_DIST_ENV=1 CUDA_VISIBLE_DEVICES=0,1,2,3 /mnt/local_storage/autoep_transformers_matrix_20260513_fixed/venv-5.8.1/bin/python -m pytest -q tests/unit/moe/test_autoep_grad_parity.py'
flock /tmp/ds_gpu_test.lock -c 'PYTEST_XDIST_WORKER=gw1 DS_DISABLE_REUSE_DIST_ENV=1 CUDA_VISIBLE_DEVICES=0,1,2,3 /mnt/local_storage/autoep_transformers_matrix_20260513_fixed/venv-5.8.1/bin/python -m pytest -q tests/unit/moe/test_autoep_unit.py'
git diff --check

Results:

• test_autoep_grad_parity.py: 1 passed
• test_autoep_unit.py: 32 passed
• git diff --check: passed