Eliminate cached queries from reproductions

[lintool]

Currently, for many reproductions involving encoders, we have two versions:

• inference on-the-fly - run the query through ONNX, PyTorch, etc. to produce the embeddings
• cached queries - caches the embeddings

The original intent was that for faster iteration, we could just use the cached queries... but:

• this slowed down reproduction tests, because we had to test both version
• over time, the embeddings drifted slightly e.g., different versions of PyTorch produced slightly different embeddings

My current thinking: there's little point in having cached queries - more to test, and they don't appear to solve an actual pain point...

So, my proposal is to remove cached queries in our reproductions.

Thoughts?

[sahel-sh]

+1, cached queries don't really test much other than faiss/search working as expected?

[lilyjge]

+1, I think it actually makes more sense reproducibility wise to make sure re-encoding the queries give the expected results, ie explicitly including query encoding to always be part of the pipeline instead of relying on pre encoded embeddings