Add data checkpoint

[floatingbigcat]

Add data checkpoint feature within one epoch, also add support for openclipb32.
We split the total train shards into small sub-shards set, we train one sub-shards set for one interval between checkpoints.
Therefore we must hold certain constrains:

• number of shards of sub-shards sets must larger than total workers(numboer of workers times world size of data parallel).
• checkpoint save way must be linear for spliting the shards evenly.

To regenerate data iterator everytime we switch to a new sub-shard sets, I feed the dataloader to our train function and rebuild the data iterator upon it everytime we hit the save iteration.

[floatingbigcat]

We use streaming dataset for supporting interleaved data input.
The solution based on webdataset will be abandoned.