ES|QL completion command constant folding

[afoucret]

Summary

This PR implements constant folding support for ESQL COMPLETION inference plans, allowing completion operations with constant prompts to be evaluated at optimization time rather than at query execution time.

Closes #136863

Technical Details

• A CompletionFunction as a internal primitive only and is not exposed to users.
• During the analysis phase in Analyzer::resolveInferencePlan where foldable Completion plans are converted to Eval nodes with CompletionFunction expressions

FROM books
| COMPLETION "Translate this text" WITH { "inference_id": "my-model" }

is internally rewritten into

FROM books
| EVAL completion=COMPLETION("Translate this text", "my-model")

• The pre-optimizer uses the InferenceFunctionEvaluator which handles the actual execution of the inference using CompletionOperator

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[elasticsearchmachine]

Hi @afoucret, I've created a changelog YAML for you.

[carlosdelest]

This LGTM 👍

I was wondering if some rules could be taken out of the Analyzer - like transforming the command into an Eval - but I don't see a special benefit out of it, and I prefer to have all the related transformations in one place 👍

Folding is straightforward and incorporated into the InferencePlan.

[ioanatia]

I added some non blocking comments that you can address before merging

[ioanatia]

can we add another test, where we have something like:

FROM books METADATA _score
| EVAL prompt = "Translate this text in", language = "French"
| COMPLETION concat(prompt, language) WITH ....

so we not only test for literals which are always foldable, but also foldable expressions that are not literals.

[ioanatia]

can we add more details here, that this is only used for optimizing the COMPLETION command when we deal with a foldable prompt with some example, that this should never be added to the function registry, since we don't have support for async functions etc.