Fix: eliminate redundant forward pass in evaluate_on_test_set

[alobroke]

What this PR does

Fixes the performance issue reported in #3.

evaluate_on_test_set() was running the model on the full test set
twice — once in its own loop and again inside validate_all().
The all_preds and all_targets collected in the first loop were
never used for metrics.

Changes made

• Added correct_pixels and total_pixels accumulators to the
  first loop
• Used extend instead of append to collect flat predictions
  and targets directly
• Replaced the redundant validate_all() call with a direct call
  to calculate_metrics() using data already collected in the loop
• Added calculate_metrics to imports from common_metrics

Result

The model now runs on the test set exactly once in all cases,
cutting evaluation time roughly in half.

Related issue

Closes #3