Duplicate data returned when use `num_workers` param (multi-processing) in Dataloader

[Albert-Ma]

`trainloader = mz.dataloader.DataLoader(
dataset=trainset,
stage='train',
num_workers=30,
callback=padding_callback
)``

Result:

It shows that model will train num_worker time, 30 in this experiment, in one epoch!

A Possible Solution: splitting workload across all workers in Dataset/ __iter__ function!