Log base CE loss (exclude aux) as train/loss

[pc0618]

Changes logging so W&B train/loss reflects base next-token CE (total minus aux), while still optimizing and logging total loss and aux separately.

• Add compute_next_token_loss_terms() to expose (base_loss, aux_loss)
• Train loops log train/base_loss, train/aux_loss, train/total_loss
• Trainer uses train/base_loss for StepInfo.loss so the default hook logs base CE as train/loss

Tests: uv run --with pytest pytest lib/levanter/tests/test_lm_model_loss.py -q

[dlwh]

this is two distinct changes, can you split