What does this PR do?

Problem

When evaluating models loaded in bfloat16 precision (e.g., Llama-3 on Ampere+ GPUs), avg_losses.cpu().numpy() fails with TypeError: Got unsupported ScalarType BFloat16 because NumPy does not support bfloat16 natively.

Solution

Explicitly cast the tensor to float32 using .float() before converting to NumPy in src/evals/metrics/utils.py.

Fixes # (issue)
N/A

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Have you gone through the contributions guide?
• Are your changes documented? Read documentation guidelines here.