Audit the onnx pathways to make them robust against >2Gb models

[dbogunowicz]

Uses the helper functions from sparsezoo to work reliably with the larger models (bigger than protobuf size, where the large data information is stored in the additional accompanying file).

Replaced all the instances where onnx model is saved with the sparsezoo helper function.
Replaced all the instances where onnx model is checked with the sparsezoo helper function.

GHA fail because we are not using the updated sparsezoo repository.
After pip installing this branch: neuralmagic/sparsezoo#308 locally, I got all green tests for pytorch and onnx test targets.