fix: initialize best_param before training loop to prevent UnboundLocalError

[Ayush10]

Summary

• Initializes best_param with the model's initial state dict before the training loop in 21 PyTorch model files
• Prevents UnboundLocalError when no epoch improves the validation score (e.g. due to NaN scores from missing data or aggressive early stopping)
• Follows the pattern already established in pytorch_gru.py, which was the only model file that handled this correctly

Root Cause

best_param was assigned only inside the if val_score > best_score: branch but used unconditionally after the loop via model.load_state_dict(best_param). When val_score is NaN (common with missing data), the comparison val_score > best_score is always False, so best_param is never assigned.

Files Changed (21 files, 1 line each)

pytorch_adarnn, pytorch_add, pytorch_alstm, pytorch_alstm_ts, pytorch_gats, pytorch_gats_ts, pytorch_gru_ts, pytorch_hist, pytorch_igmtf, pytorch_krnn, pytorch_localformer, pytorch_localformer_ts, pytorch_lstm, pytorch_lstm_ts, pytorch_sandwich, pytorch_sfm, pytorch_tabnet, pytorch_tcn, pytorch_tcn_ts, pytorch_transformer, pytorch_transformer_ts

Test plan

• Verified each file has exactly 1 line added with the correct model variable name
• Verified pytorch_gru.py (already fixed), pytorch_general_nn.py (uses step==0 guard), pytorch_tra.py (different pattern), and pytorch_tcts.py (loads from file) were not modified
• CI pipeline passes

Fixes #1794