Added support for latest pytorch-lightning, albumentation and fixed multiple issues with DDP training, images sizes normalization and focal loss

[ExtReMLapin]

• Added support for pytorch lightning and it's latest version,
  • updated few methods names and added options like ddp_find_unused_parameters_false
• Fixed crash with batch size != 1 with images of different sizes using padding (on square image) and --max_image_size which defaults to 512
  • i could have added smarter padding instead of box of size 512x512
• implemented focal loss using --use_focal_loss
• fixed multiple errors with ddp
  • crashes and errors linked to tensors being split on multiple devices including CPU
  • slow start up because of non recycled dataloaders (recreated on each epoch)
  • images not uploaded to wandb because of missing ENV variable
• implemented better schedulers

Please note that when increasing image size you obviously need to increase points size (--heatmap_sigma) and other evaluation metrics like

• --minimal_keypoint_extraction_pixel_distance
• --maximal_gt_keypoint_pixel_distances

[ExtReMLapin]

Pr is still draft because things need to be cleaned up a bit like the gitignore and my very aggresive albumentations options