Move dataloader initialize_object to factory methods

[hanlint]

This PR continues our effort to remove logic from initialze_object in preparation of deprecating yahp from the codebase. We move most of the logic into factory methods, e.g.:

def build_ad20k_dataloader (...) -> DataSpec
   ...

def build_synthetic_ade20k_dataloader(..) -> DataSpec
  ...

TODO:

• ADE20k
• ImageNet
• C4 @growlix
• BRATS
• CIFAR @dblalock
• COCO @A-Jacobson
• MNIST
• LMDataset @growlix

Need some help to complete the rest!

[knighton]

ready for review or wait for all checks?

[hanlint]

@knighton wait for all checks, thank you :)

[dblalock]

Added CIFAR-10 dataloader factories. Also added tests for CIFAR10 and MNIST based on test_c4.py. There's a bunch of duplicate code across {CIFAR,MNIST} factory funcs and tests right now. Might be worth separating some code out into shared utilities once we have something working for all the datasets / know exactly what we'd like to DRY out.

Test code is especially redundant; only holding off on adding shared utils since maybe there's already a plan to replace test_dataset_registry.py.

[hanlint]

Yeah we can DRY these out later.

@coryMosaicML since we are dropping BRATS, can you help test the imagenet/ade20k ones to make sure nothing broke with this PR? I think just kicking off a run and making sure it successfully starts training would be sufficient.

[hanlint]

@eracah @knighton this is ready for review!

[eracah]

Great work, @hanlint! This is super important for de-YAHPing because the dataset Hparams is where a lot of the "black magic" of hparams lives. Thanks for doing this. Looks pretty good to me, but I'll defer to James, who probably knows this stuff better than me

[A-Jacobson]

noted what i think is a missing param in ade20k_hparams. otherwise lgtm! good stuff.