Fix eval loss normalization in BPB evaluation

[art-test-stack]

Summary

• remove the extra batch-size division from the debug loss returned by compute_bpb
• add a regression test using multiple batches and a batch size greater than one
• keep the BPB calculation unchanged

Cause

loss2d.mean() already averages the per-token loss over both batch and sequence dimensions. The accumulated batch means were subsequently divided by steps * batch_size, so eval/loss was smaller than the actual mean cross-entropy by exactly the per-rank batch size.

The correct denominator for the existing accumulator is world_size * steps.

Test

The regression test evaluates a constant loss of 2.5 with batch size 4 and two evaluation steps. The old implementation reports 0.625; the corrected implementation reports 2.5.