evaluate_model() returns placeholder None values

[SahilKumar75]

trainer/evaluate.py loads perplexity and BLEU metrics but the evaluation loop is not implemented — it always returns {"perplexity": None, "bleu": None}.

[SahilKumar75]

Resolved by #33 — trainer/evaluate.py now computes real perplexity using model.eval() + torch.no_grad() on a 20% held-out sample. Result is published to Redis job:{id}:eval and displayed in the wizard results page. BLEU is intentionally omitted — instruction-following datasets have no reference outputs.