feat: support back again Python 3.6 modules

[IgnacioHeredia]

This PR is to backport deepaas-cli fixes to Python 3.6, in order to allow old modules from the marketplace to be deployed with OSCAR.

I have started form the latest 2.5.2 version (branch releases/2.x) and done the following changes:

• removed groups from TOML (only used for testing purposes), because not compatible with the Poetry version available in 3.6 (1.1.15) (ref, ref)
• downgraded dependencies in TOML until reaching Python 3.6-compatible versions
• commented importlib.metadata (not available in 3.6-compatible version of importlib)

⚠️ Note: I haven't modified Github workflows because I understand this will be a one-time backport.

Tested: This has been tested with the ai4os-image-classification package (Python 3.6.8), with:

• model metadata
• train
• predict
• deepaas-cli

root@b81b8cb2f65f:/srv/deepaas# deepaas-cli predict --files "/srv/ai4os-image-classification-tf/data/samples/labrador.jpeg" 

[...]

1/1 [==============================] - 1s 740ms/step
{'status': 'OK', 'labels': ['golden_retriever', 'Irish_setter', 'gibbon', 'Saluki', 'Chihuahua'], 'probabilities': [0.61965012550354, 0.09341743588447571, 0.04869064316153526, 0.04133171588182449, 0.028836732730269432]}

[sonarcloud]

Quality Gate failed

Failed conditions
D Maintainability Rating on New Code (required ≥ A)

See analysis details on SonarCloud

Catch issues before they fail your Quality Gate with our IDE extension SonarLint