Fix five pipeline-correctness bugs

[evasnow1992]

Summary

Five correctness fixes found during a focused review of the training, finetuning, and inference pipelines. Every fix was verified in-container (torch 2.9.1) before committing. No refactors, no behavior changes beyond the bug being corrected.

Fixes (one per commit)

1. fix(predict): filter invalid SMILES before featurization task/predict.py built valid_indices as list(range(len(test_data))), i.e. it never filtered — the None-insertion/realignment code below it was dead. Any unparseable or empty SMILES (or a zero-heavy-atom molecule) in a prediction CSV would crash featurization instead of being skipped. Now validity is checked with RDKit (MolFromSmiles, ≥1 heavy atom) exactly as the training path does, so invalid rows are dropped and predictions align to the surviving molecules.

2. fix(finetune): zero MTLLoss.log_sigma grad each step (--use_mtl_loss) MTLLoss.log_sigma is a standalone nn.Parameter appended directly to the optimizer (param_groups[1]); it is not a submodule of model, so the per-batch model.zero_grad() never cleared its gradient. The gradient accumulated across every batch, corrupting the learned task-uncertainty weights. Added mtl_loss.zero_grad() alongside model.zero_grad(). Verified with the real MTLLoss: log_sigma.grad went 1→2→3 across steps before the fix, constant per-step after.

3. fix(eval): use BCE on probabilities for classification eval loss In eval the model applies sigmoid to classification outputs, so the eval predictions are already probabilities. The eval-loss path then passed them to BCEWithLogitsLoss, which sigmoids a second time — a double sigmoid that produced a wrong loss (off by up to ~7x on confident predictions). Now the classification branch uses plain BCE on the probabilities; regression is untouched. Verified the corrected loss matches BCEWithLogits on the raw logits to within 1e-7.

4. fix(eval): keep per-task alignment when a task has no valid targets In evaluate_predictions, a task with zero valid targets hit continue without appending to results, so results came out shorter than num_tasks and every later task's metric silently shifted down one index. Now appends float('nan') before continue, matching the all-0s/all-1s branch directly above it.

5. fix(finetune): don't double-wrap DataLoader for ensemble_size > 1 The training loop assigned the wrapped loader back to train_data, the same name holding the underlying dataset. The first ensemble model was fine, but the second iteration then did DataLoader(DataLoader(...)), which crashes with "'DataLoader' object is not subscriptable". Use a separate train_loader local and keep train_data as the dataset. Verified with a 2-iteration repro (old crashes on iter 1, fixed runs both).

Behavior impact (what changes)

• Predict (Docker container build error #1): CSVs with invalid SMILES now run to completion instead of crashing; output rows correspond to the valid molecules only.
• Finetune with --use_mtl_loss (make cuik-molmaker optional #2): the multi-task uncertainty weights now train correctly. Any prior --use_mtl_loss run had corrupted task weighting. --use_mtl_loss is off by default.
• Classification eval/val loss value (Molecules clustering code (PCA and K-means) for training datasets: #3): now correct. Metric-based (AUC) selection and predictions are unchanged; loss-based checkpoint selection and HPO for classification now see the correct loss.
• Multi-task eval with partially-labeled splits (Fix TypeError when dynamic_depth="truncnorm" by using rvs() instead of rvs(1) #4): per-task metrics are now correctly aligned. Single-task and fully-labeled multi-task runs unaffected.
• Ensemble finetune (Disallow inconsistent usage of finetune and predict workflows #5): --ensemble_size > 1 now works; it was previously broken and could not have completed.