feat: combine embeddings

[JohannesMessner]

Right now it is painful to combine embeddings from different nesting levels and set them at the top level, especially until https://github.com/jina-ai/docarray/issues/461 is solved.

So this should do the following: It takes an access path (or list of access paths?) (e.g. @.[image, main_text]), and combines the embeddings in those docs using one of 'sum', 'mean', 'concat', or a provided model.

It should handle the following cases:

• [x] numpy embedding / predefined combiner
  • [x] test
• [x] torch embedding / predefined combiner
  • [x] test
• [ ] tf embedding / predefined combiner
  • [ ] test
• [ ] paddle embedding / predefined combiner
  • [ ] test

---

• [x] torch embedding model
  • [ ] test
• [ ] tf embedding model
  • [ ] test
• [ ] paddle embedding model
  • [ ] test
• [ ] onnx embedding model
  • [ ] test

---

• [x] numpy embedding / callable
  • [ ] test
• [x] torch embedding / callable
  • [ ] test
• [x] tf embedding / callable
  • [ ] test
• [x] paddle embedding / callable
  • [ ] test

Other ToDos:

• [ ] docs
• [ ] rename the method: fuse_embeddings()?
• [ ] refactor. Currently the code is quite ugly
• [ ] refactor examples in the docs

Possible follow-up PRs:

• [ ] enable a flag uniform_nesting which tells us that every doc in the da has the same number of relevant chunks. This would allow us to vectorize the combine operation
• [ ] implement to_numpy
• [ ] implement flag to discard chunk embeddings after root embedding has been set

Closes #512