multi-GPU training breaks for main_pretrain.py

[alkalait]

Multi-GPU training (say, with devices 0, 1)

python main_pretrain.py \
  --data_dir datasets/seco_1m --data_mode seco \
  --base_encoder resnet18 --gpus 2

suffers from the same issue raised in Lightning-AI/pytorch-lightning#1506

The way to deal with it is to set --replace_sampler_ddp False, like so:

python main_pretrain.py \
  --data_dir datasets/seco_1m --data_mode seco \
  --base_encoder resnet18 --gpus 2 --replace_sampler_ddp False