Make reference scripts compatible with submitit

[datumbox]

This PR makes the train.py reference scripts of Classification, Detection and Segmentation compatible with SubmitIt.

We use a modified version of the script from DETR. The actual runner script will not be added in the repo but you can fetch the compatible version from here.

To ensure BC, we check that all possible runners remain functional by running the following commands:

# Classification
PYTHONPATH=$PYTHONPATH:`pwd` python -u run_with_submitit.py --timeout 3000 --ngpus 8 --nodes 2 --model mobilenet_v3_large --test-only --pretrained
python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --model mobilenet_v3_large --test-only --pretrained
torchrun --nproc_per_node=2 train.py --model mobilenet_v3_large --test-only --pretrained
sbatch launch_job.sh --model mobilenet_v3_large --test-only --pretrained

# Detection
PYTHONPATH=$PYTHONPATH:`pwd` python -u run_with_submitit.py --timeout 3000 --ngpus 8 --nodes 2 --dataset coco --model ssd300_vgg16 --pretrained --test-only
python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --dataset coco --model ssd300_vgg16 --pretrained --test-only
torchrun --nproc_per_node=2 train.py --dataset coco --model ssd300_vgg16 --pretrained --test-only
sbatch launch_job.sh --dataset coco --model ssd300_vgg16 --pretrained --test-only

# Segmentation
PYTHONPATH=$PYTHONPATH:`pwd` python -u run_with_submitit.py --timeout 3000 --ngpus 8 --nodes 2 --dataset coco --model lraspp_mobilenet_v3_large --test-only --pretrained
python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --dataset coco --model lraspp_mobilenet_v3_large --test-only --pretrained
torchrun --nproc_per_node=2 train.py --dataset coco --model lraspp_mobilenet_v3_large --test-only --pretrained
sbatch launch_job.sh --dataset coco --model lraspp_mobilenet_v3_large --test-only --pretrained

All return the same expected results.

[fmassa]

Thanks!