Initial NaFlex ViT model and training support

[rwightman]

Working:

• 'flex' ViT w/ NaFlex position embedding resize, pre-patched input, attention padding masks
• Single node train.py works with a custom naflex data-pipeline via a dataset wrapper that handles random seq-len & batch-size selection, constrains images to seq-len while keeping aspect ratio (with randomizations)
• A much faster patch embed kernel resample, torch only, can be used in forward()

Not tested / not completed:

• distributed training not tested, dataset wrapper needs more verification
• dataset wrapper for iterable datasets (wds, tfds, iterable hfds) needs to be added
• more model definitions
• weight loading / translation for existing vits
• SigLip-2 NaFlex vision encoder weight port
• Integration of naflex data pipeline components into OpenCLIP
• Add randomization of the patch_size along with seq_len

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.