arroy for vectors

[ricopinazo]

What changes were proposed in this pull request?

Why are the changes needed?

The old approach of defining custom dcument search queries in python had two problems:

1. It used the python host as the runtime, leading to potential deadlocks etc
2. It used to load all the vectorised graphs in the server, which didn't scale

Does this PR introduce any user-facing change? If yes is this documented?

How was this patch tested?

Are there any further changes required?

• remove heed from deps
• try to remove all the duplication for tanslating between u32 and entity id
• double-check having max lmdb size set to 1TB doesnt cause the disk usage to grow for small graphs
• have a separate module for the db abstraction
• keep adding benchmarks, use different query for each call in the current one
• stop loading all graphs for the global plugins
• Reuse EntityRef for DocumentEntity
• re-enable save_embeddings function
• move GqlDocument under graph folder
• add window parameters to vertor selection and vectorised graph functions in GraphQL
• reiview if Im actually using async_stream crate and the others
• double-check Ctrl+C still works
• decide proper size for the heed env
• test the miss rate in a real-world scenario: 0.02% misses in a 100k dataset
• remove all unwraps from cache.rs