any-llm batch API requires all providers to normalize to OpenAI Batch/BatchResult types with provider-specific status mappings

The any-llm batch completion architecture requires every provider to convert native batch job representations to OpenAI's Batch type (from openai.types.Batch) and results to BatchResult (custom dataclass with ChatCompletion or BatchResultError items). Each provider needs: a status map from native statuses to OpenAI statuses (validating/in_progress/completed/failed/cancelling/cancelled/expired), a conversion function for the native job to Batch, and a results parser. The five methods to implement are _acreate_batch, _aretrieve_batch, _acancel_batch, _alist_batches, and _aretrieve_batch_results. Providers like Anthropic use inline requests (read local JSONL, convert each entry), while providers like Bedrock require cloud storage URIs (S3/GCS). Unknown statuses should log a warning and default to in_progress. Verified 2026-05 against any-llm.

When implementing batch support for a new any-llm provider: (1) set SUPPORTS_BATCH=True, (2) create a status map dict, (3) implement all 5 _a* batch methods with @override, (4) normalize all returns to Batch/BatchResult types, (5) raise BatchNotCompleteError for results on incomplete batches, (6) raise InvalidRequestError for missing required kwargs. Check existing implementations (Anthropic for inline pattern, Bedrock for cloud-storage pattern) as references.