More intensive tests for fit/inference

[Joao-L-S-Almeida]

The problem with the tests using fit and inference pipelines is that not all the architectures have versions available on HuggingFace or can be slow to download during user-side tests. In order to circumvent that I suggest:

• To automatically create manufactured (fake) checkpoints to be used together with template YAML configs.
• These checkpoints store randomly generated weights, just used to test the execution via command line.
• The checkpoints generation and command line tests can be automatically executed via GitHub Actions.
• The manufactured checkpoints can be used to perform equally manufactured fine-tuning/prediction processes for a few epochs, just to guarantee the toolkit general usability for those applications.

Any suggestion @CarlosGomes98 and @biancazadrozny ?

[Joao-L-S-Almeida]

Please check the PR attached to this issue in order to see if the proposal was covered.