Default to single-process batch inference

[mdekstrand]

Related to #841 and #921, for 2026 we should look at defaulting batch recommendation / prediction to one thread / process, especially when Ray is not available.