Simple addition to enable bidirectional neighbor loader benchmarking.
Results show a relative drop in performance moving from directional to bidirectional (example below) maybe due to the fact that, currently, bidirectional sampling is implemented as a post-processing of regular sampling #7200

With directional sampling:
Dataset: products
Training sampling with [10, 5] neighbors
100%|█████████████████████████████████193/193 [00:03<00:00, 48.96it/s]
100%|█████████████████████████████████ 193/193 [00:03<00:00, 49.59it/s]
100%|█████████████████████████████████ 193/193 [00:03<00:00, 49.01it/s]
batch size=1024, iterations=579, runtimes=[3.943, 3.892, 3.939], average runtime=3.925

while with bidirectional sampling:
Dataset: products
Training sampling with [10, 5] neighbors
100%|██████████████████████████████████ 193/193 [00:04<00:00, 42.22it/s]
100%|██████████████████████████████████193/193 [00:04<00:00, 43.18it/s]
100%|██████████████████████████████████193/193 [00:04<00:00, 42.56it/s]
batch size=1024, iterations=579, runtimes=[4.572, 4.47, 4.535], average runtime=4.526