Add GNN support to bb-search benchmark pipeline

Fix --gpu-data OOM: in-place normalization, staged GPU transfer

Add --gpu-data flag to store training data on GPU for faster training

Fix Colab T4 OOM: CPU-resident data, batched validation, --max-samples

Fix Colab double-cd bug: use absolute path for repo dir

Update Colab notebook for human game GNN training

GNN architecture, human game feature extractor, inference integration

Batch ranking loss forward pass + increase batch size for GPU

Add ranking loss training to fix MCTS value head move discrimination

Update Colab notebook for exp #23 asymmetric loss training

Add asymmetric loss + negative oversampling for value head training

AZ training memory/resume fixes, MCTS hybrid experiments, Colab GPU training

WIP: AZ split-body net, ExIt self-play pipeline, bridge trade strategy, misc fixes

Only disable OMP threads when --workers is passed

Fix fork deadlock: disable OMP/MKL threads before torch import