Fix zero/division safety gaps in utility and inference paths

[harshang03]

Describe your changes

• Added a shared non-zero divisor validator and wired it into group divisibility checks and inference ceil_div.
• Added strict steps_per_output validation in ThroughputTimer so invalid values fail early instead of triggering modulo-by-zero at runtime.
• Hardened HPU FP8 dequantization to reject zero or non-finite scales before inverse-scale computation.
• Added targeted regression tests for groups, timer, inference utils, and HPU quantizer scale validation.

Screenshot or video (only for visual changes)

• N/A

GitHub Issue Link (if applicable)

• [BUG] Multiple missing zero-guards cause ZeroDivisionError / non-finite values across DeepSpeed (4 locations) #7838

Testing Plan

• Explanation of why no additional tests are needed:
  • Added focused unit tests that directly cover each reported failure mode and guard path.
• Unit Tests (JS and/or Python):
  • ./.venv/bin/python -m pytest tests/unit/utils/test_groups.py tests/unit/utils/test_timer.py tests/unit/inference/test_inference_utils.py tests/unit/ops/fp_quantizer/test_fp_quantizer_scale_validation.py
• E2E Tests:
  • Not run (change is utility-level and covered by unit tests).
• Any manual testing needed?:
  • No.

---

Contribution License Agreement
By submitting this pull request you agree that all contributions to this project are made under the Apache 2.0 license.