AbstractGraph refactor

[f-aguzzi]

Is your feature request related to a problem? Please describe.
The code:

• is messy (long module, irregular implementation)
• has known bugs (some custom embeddings do not work even if implemented, see Not match Gemini embeddings model #496)
• is hard to read (irregular formatting)
• is difficult to understand for newcomers and contributors

Describe the solution you'd like
Ideally, the module should be rewritten from scratch at some point. In the meanwhile, all dead code should be removed (for example, the RAG node has been deprecated in all graphs, so the sections about embeddings have become useless).
A better use of LangChain and of its latest features should remove some of the complexity.

Describe alternatives you've considered
Possible alternatives will be considered and explained in the follow-ups to this issue.

[f-aguzzi]

The main tasks are:

• clean up the code and enforce correct formatting
• remove dead imports
• transplant the embedding initialization blocks into the RAGNode

[f-aguzzi]

The process should fix the following issues too:

• 'Batchsize' config attribute does not change #413 (batch_size kwarg)
• Not match Gemini embeddings model #496 (Gemini embeddings not matching pattern)
• does it support custom embeddings #460 (choice of custom embeddings)

[f-aguzzi]

Also this:

• result is empty for any url from domain http://www.mckinsey.com #501 (Azure-hosted GPT models get incorrectly treated as OpenAI)