[compute-ai-embeddings] Rework support for Hugging Face local models for embeddings

[eolivelli]

Summary:

• Change Hugging face "modelUrl" parameter to "model-url"
• Use the "provider" value from the Hugging Face resource instead of picking it from the "agent" (now you can have multiple HF resources), provider can be "api" or "local"
• Add some tweaks to make HF local computation work again (it used to work but after a couple of refactors we broke the configuration)
• Add one test that uses HF locally
• Add a sample application that uses an HF embedding model (intfloat/multilingual-e5-small)
• Update the example secrets file to allow to configure HF provider and models with ENV variables (at least for the example application)
• Upgrade DLJ an PyTorch (not strictly needed but as we are touching this part I preferred to pick up the latest and greatest)

Please note that the integration tests shows that when the test is running on CI the precision of the embeddings seems "better", at least it emits one more digit