Multiple nodes training

[skpig]

1. Add Multi_Node_Training in image_classification.
2. Write a tutorial about multi-node training in README.md
3. Use ViT model as an example.

[xperzy]

BTW, in the main_multi_gpu.py, please also check the logging/model saving scheme such as

PaddleViT/image_classification/Multi_Node_Training/main_multi_node.py

Line 361 in c6be596

if local_rank == 0:

Current code does not consider the case where world_size > 1.

[skpig]

I just check the details of spawn() function. For example, we have 2 hosts with 2 processes running on each host. Then the
local_rank = dist.get_rank() will return 0, 1, 2, 3 respectively. I guess the original code works fine?