Normalize cloned grad_absmax metrics for structured progress payloads

[Copilot]

Structured progress/webhook metrics were inconsistent between grad clipping modes: the "norm" path serialized cloned values to floats, while the "value" path could emit a Tensor for grad_absmax. This updates the value-clipping path to normalize cloned metrics the same way.

• Metric serialization

  • Update Trainer._update_grad_metrics() so clone_norm_value=True converts grad_absmax to a detached float in the "value" clipping path.
  • Preserve existing behavior for accelerator logging paths that still use the original tensor object when cloning is not requested.

• Focused coverage

  • Add a unit test for cloned grad_absmax serialization in the value-clipping path.
  • Update the progress-metrics expectation to assert the serialized float form for structured payload generation.

• Behavioral example

  # Before: could place a Tensor into structured metrics
  trainer._update_grad_metrics(metrics, clone_norm_value=True)
  metrics["grad_absmax"]  # torch.tensor(...)
  
  # After: normalized for JSON-serializable payloads
  trainer._update_grad_metrics(metrics, clone_norm_value=True)
  metrics["grad_absmax"]  # 1.2