Fix training/validation metrics scaling and validation loss aggregation

[YonghaoZhao722]

This PR fixes two issues in training and validation metrics logging that affect comparability and correctness.

Problems fixed

1. Training loss / metric scaled with batch size

In SamTrainer._compute_loss(), per-sample mask_loss and iou_regression_loss were accumulated across the batch without normalization.
As a result, increasing batch_size inflated the reported loss and metric values even when model quality was unchanged, making runs hard to compare.

2. Validation mask_loss / iou_loss were not epoch averages

During validation, validation/metric and validation/loss were averaged over len(val_loader), but validation/mask_loss and validation/iou_loss effectively reflected only the last validation batch rather than epoch-level averages.

Changes made

• Normalize mask_loss and iou_regression_loss by batch_size inside _compute_loss() so training metrics are batch-size invariant.
• Accumulate and log epoch-averaged validation/mask_loss and validation/iou_loss.
• Apply the same aggregation logic to the joint trainer for consistency.

Why this matters

• Makes training curves comparable across different batch sizes, which is important for scaling experiments.
• Ensures TensorBoard validation scalars accurately represent epoch-level averages.

This change does not affect model behavior or optimization—only how metrics are computed and reported.

[anwai98]

Looks good to me. I'll merge this in a few minutes. Thanks for your contribution!