Add automated tests for checking correct state preservation in models

[olgavrou]

Short description

We keep getting bug reports (usually coming from python) where the predictions between a trained vw instance differ if you save the model and then load it (without any training in between). i.e.:

predictions from:

1. initialize vw instance
2. train
3. predict

differs from

1. initialize vw instance
2. train
3. save model (not referring to save_resume here)
4. load model
5. predict

Possible solution/implementation details

This could be automated using the python runtests infra to (for each test in RunTests):

1. create vw instance from a run_test test
2. train it
3. predict -> keep predictions 1
4. save
5. load
6. predict -> keep predictions 2
7. compare predictions 1 with predictions 2

[lalo]

similar to flatbuffer automated tests https://github.com/VowpalWabbit/vowpal_wabbit/blob/2d394ce7bdb02cc3f5bb08c0031565e0e4ea48dc/test/run_tests.py#L19

[michiboo]

@olgavrou Hi, I would like to try this issue if it is not done yet.