chunkable datasets and dataloaders

Nobody has claimed this yet.

### Description & Motivation

Current large model training requires a huge number of traning samples, so that the traditional Mapped dataloaders are failed to load the training data because of limited memory. so I think there can be a chunkable datasets and dataloaders available. so that we can use the Mapped logic to load the first subset of training data in the dataset and dataloader 1 for training and the loading the second subset of training data for prepare.

### Pitch

I had seen the CombinedLoader in lighting, and I did not find too much documents and examples about it, It seems that it can not resolve the requirement about **"loading large traning data in memory chunk by chunk"**. If there had been some resolution, please give me some help, thanks!

### Alternatives

_No response_

### Additional context

_No response_

cc @lantiga @borda @tchaton

Start by reading the CombinedLoader documentation and examples, then compare its behavior with the mapped dataloaders described in the issue. The work is complete when a documented approach supports loading large training data chunk by chunk without requiring the full dataset in memory.

Written by the indexing model from the issue text.