fix docs examples for ForwardBackwardOutput metrics access

[lokashrinav]

Summary

This updates TrainingClient documentation examples to reflect the actual ForwardBackwardOutput shape.

ForwardBackwardOutput does not expose a .loss attribute. Accessing result.loss raises an AttributeError.
For built-in cross-entropy examples, loss is available under:

result.metrics["loss:sum"]

For custom loss examples, metrics are user-defined, so the docs now show printing the full result.metrics map instead of assuming a loss:sum key always exists.

What changed

• Replaced .loss usage in standard forward/forward_backward examples with metrics["loss:sum"]
• Updated custom loss examples to print result.metrics (without assuming loss:sum)

Why

Running the documented examples as written currently fails with:

AttributeError: 'ForwardBackwardOutput' object has no attribute 'loss'

This PR aligns the docs with the real API surface and avoids misleading assumptions in custom-loss workflows.

Reproduction

fwdbwd_future = await training_client.forward_backward_async(data=[datum], loss_fn="cross_entropy")
fwdbwd_result = await fwdbwd_future.result_async()
print(fwdbwd_result.loss)  # AttributeError