fix(exporters/tasks): support local timm model paths in ONNX export

[Anai-Guo]

Problem

When exporting locally fine-tuned timm models to ONNX, the export fails with:

Repository Not Found for url: https://huggingface.co/my-local-model/resolve/main/config.json

Root cause: get_model_from_task unconditionally prepends hf_hub: to the model path when loading timm models (line 1177 in optimum/exporters/tasks.py):

# Before fix — always uses HF Hub even for local paths
model = model_class(f"hf_hub:{model_name_or_path}", pretrained=True, exportable=True)

This forces timm's create_model to treat the path as an HF Hub model ID, breaking local directory paths.

Fix

Check if model_name_or_path is a local directory. If so, pass it directly to create_model; otherwise use the hf_hub: prefix for Hub models.

# After fix
if os.path.isdir(model_name_or_path):
    model = model_class(model_name_or_path, pretrained=True, exportable=True)
else:
    model = model_class(f"hf_hub:{model_name_or_path}", pretrained=True, exportable=True)

Testing

# Local timm model (should now work)
optimum-cli export onnx --model ./my-local-timm-model --task image-classification onnx_output/

# HF Hub timm model (should still work)
optimum-cli export onnx --model timm/efficientnet_b0.ra_in1k --task image-classification onnx_output/

Fixes #2423