[Not for merge] Support multi-machine training

[zhu-han]

Example of multi-machine training:

Suppose we have 2 machines, each has 8 gpus,

• The command in the first machine:

python3 -m zipvoice.bin.train_zipvoice \
    --world-size 16 \
    --local-rank-start 0 \ 
    --local-world-size 8 \ 
    --master-addr xxx.xxx.x.xxx \ # Master node ip address
...

• The command in the second machine:

python3 -m zipvoice.bin.train_zipvoice \
   --world-size 16 \
   --local-rank-start 8 \ 
   --local-world-size 8 \ 
   --master-addr xxx.xxx.x.xxx \ # Master node ip address (same with the first machine)
...