Fixed test/preparation.jl line 68 where the embedding response was
incorrectly formatted as a single dictionary with a 2D matrix instead
of separate dictionaries with 1D vectors.

This was causing CI failures after upgrading to PromptingTools v0.84+,
which has stricter validation of API response formats.

Changed from:
- Dict(:embedding => ones(128, 2)) - single dict with 128x2 matrix

To:
- [Dict(:embedding => ones(Float32, 128)),
   Dict(:embedding => ones(Float32, 128))] - two dicts with 128-dim vectors

This matches the correct OpenAI API format where each embedding is
returned as a separate object in the data array.