fix: remove dead gradient norm fallback

[LOGO127]

Summary

Fixes #8371.

set_norm_for_param_grad_in_gpu previously fell back to param.grad when the
configured gradient attribute was None. That fallback is invalid for the
grad_accum path because the source gradient has already been moved and
cleared. The method now uses the selected attribute directly and reports a
clear assertion if the required gradient is missing.

Tests

• pre-commit run --files deepspeed/runtime/zero/stage_1_and_2.py tests/unit/v1/zero/test_zero.py
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 py -3 -m pytest -q tests/unit/v1/zero/test_zero.py::TestSetNormForParamGradInGPU (3 passed)
• py -3 -m compileall -q deepspeed/runtime/zero/stage_1_and_2.py tests/unit/v1/zero/test_zero.py
• git diff --check

The regression test covers both configured gradient attributes and verifies
that the stale param.grad fallback is not used.

AI assistance was used for drafting and local validation; I reviewed the
implementation and take responsibility for the submitted changes.

[LOGO127]

The latest remote validation completed successfully: modal-torch-latest / collect tests, modal-torch-latest / DeepSpeedAI CI, and DCO all passed. The focused CPU regression tests and local pre-commit checks also pass; this PR is ready for maintainer review.

[LOGO127]

Gentle follow-up on #8371: the focused CPU regression, pre-commit checks, DCO, and the remote modal-torch-latest validation remain green. The PR is still mergeable and ready for maintainer review. Please let me know if any additional coverage or adjustment is preferred.