[CLIP]: Sparseml onnx export pathways for open_clip models

[dsikka]

Summary:

• The current onnx export functionality using OpenClip requires torch nightly and opset = 14
• This PR provides an example of how to use sparseml's onnx export functionality to export the visual and text branches of the CLIP model, using OpenClip
• OpenClip wraps everything in a torch.nn.Module object for the visual branch (ie everything required for the forward pass) but does not do this for the text branch. For the sake of completeness, the TextModel class was added (see clip_models.py) which wraps the appropriate steps needed in the forward pass. This is then exported
• Furthermore, for captioning, OpenClip provides thee CoCa models, which requires an additional text-decoder and also uses a separate interface compared to the other models. As such, there is a special is_coca case provided in the scripts

Test Plan:

• Ran this locally (requires openclip and torch nightly to be installed) and created a convnext_onnx directory with the clip_text.onnx and clip_visual.onnx exports

 python clip_onnx_export.py --model convnext_base_w_320 --pretrained laion_aesthetic_s13b_b82k --export-path convnext_onnx

• Also ran a CoCa-specific command which created the default clip_onnx directory with the visual and text exports, along with the text decoder export clip_text_decoder.onnx

python clip_onnx_export.py --model coca_ViT-L-14 --pretrained laion2b_s13b_b90k

[bfineran]

LGTM pending comments

[rahul-tuli]

LGTM!