Dataloader updates and streamlining

[daviswer]

A number of cleanup/streamlining changes with no behavioral impact at the training level. This is mostly preparation for landing in torchdata and upcoming features, particularly support for n_workers > 1.

• Modify instantiation so that all path- and rank-dependent setup in a layer is deferred to a new recursive setup fn, which executes immediately before saving/loading/stepping. This allows for modification of rank/path after instantiation.
• SamplingDataset and ScalableShardDataset are now implemented as proper _WrapperDatasets (though they still must stack in the same order and location). This reflects the intended modular use of _WrapperDatasets, rather than treating these as edge cases of _StatefulDataset due to their containing multiple sub-iterators. No more pass-through args!
• Remove Weird_Separated_Camel_Case naming convention in favor of ProperClassNaming
• Update unit tests to reflect above changes
• Update dataloader builder to reflect the more modular sampling and scalable datasets
• Update PreloadBufferDataset so that when the buffer is too large (i.e. rescaling to a smaller number of workers) it will shrink down to the desired size over time, rather than staying oversized.

Again, no user-facing behavior is changed, these are just preparatory updates for further additions. Backward compatibility with older checkpoints is maintained, and training code runs unaffected.

[lchu6]

LG.

regression test passed.