Allow to specify loads for the embeddings

It does not seem the current dsl and/or embedding_model thing has a structured way to define loads.
In ecommerce you typically want to pull from at least 3-4 resources to make a relevant vector for a given product for example.

Also would be cool if you could define the prompt with an `expr`

Start by examining the current DSL and embedding_model APIs to understand how embedding inputs and prompts are defined. The work is complete when users can specify structured loads for embedding resources and define the prompt with an expr, with the supported configuration and behavior documented or covered by the project’s existing checks.

Written by the indexing model from the issue text.