More resilient eval (#32257)

Bubbles up rate limit information so that we can retry after a certain
duration if needed higher up in the stack.

Also caps the number of concurrent evals running at once to also help.

Release Notes:

- N/A

crates/language_models/src/provider/mistral.rs

@@ -364,6 +364,7 @@ impl LanguageModel for MistralLanguageModel {
         'static,
         Result<
             BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
+            LanguageModelCompletionError,
         >,
     > {
         let request = into_mistral(