Reject non-finite fp16 loss_scale across config and ZeRO paths

[harshang03]

Describe your changes

• Added centralized loss-scale validators for finite/positive numeric constraints used by runtime configuration and optimizer paths.
• Enforced fp16 config validation for loss_scale, loss_scale_window, hysteresis, and min_loss_scale so invalid values fail during config parsing.
• Hardened LossScaleConfig and CreateLossScaler to reject invalid static/dynamic loss-scale arguments even if values are injected outside config parsing.
• Added validation to ZeRO stage 1/2 and stage 3 loss-scale override/setter methods to block non-finite/non-positive runtime overrides.
• Added targeted unit tests for config parsing, loss-scaler validation, and ZeRO override/setter guard behavior.

Screenshot or video (only for visual changes)

• N/A

GitHub Issue Link (if applicable)

• [Bug] Missing input validation for fp16 loss_scale accepts float('inf') #7852

Testing Plan

• Explanation of why no additional tests are needed:
  • This change is fully covered with focused unit tests at config, scaler, and ZeRO runtime guard layers.
• Unit Tests (JS and/or Python):
  • python -m pytest tests/unit/runtime/test_ds_config_model.py tests/unit/runtime/half_precision/test_loss_scale_validation.py
• E2E Tests:
  • Not run (validation and guard logic change).
• 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.