Add Batch API support (Anthropic/OpenAI) for ~50% cheaper async translation

[jaylann]

Motivation

Provider Batch APIs (Anthropic Message Batches, OpenAI Batch) process requests asynchronously at roughly 50% of the standard per-token cost, which is ideal for large localization jobs where a 24h turnaround is acceptable. This was requested alongside the OpenRouter work but deferred.

Why it's non-trivial

pydantic-ai (now on v1.x in this repo) has no batch-API abstraction. The current translation path uses pydantic_ai.Agent(...).run_sync() with structured output (output_type=TranslationResult). Batch mode cannot reuse that and would need a separate code path:

• Build raw request payloads per provider (native anthropic / openai SDKs), replicating the structured-output contract (tool/response_format) that pydantic-ai currently handles.
• Submit the batch, then poll for completion (minutes to ~24h) and retrieve/parse results.
• Persist job state (batch IDs ↔ file/lang/keys) so a run can be resumed after the process exits, since results aren't available synchronously.
• Map results back into the xcstrings entries and reconcile partial failures.

Proposed shape

• Opt-in --batch flag on translate.
• Implement Anthropic first, then OpenAI. OpenRouter/Gemini batch are out of scope initially.
• Sub-commands or flags to submit vs. poll/collect a previously submitted batch (job-state file in the working dir).

References

• Anthropic Message Batches API
• OpenAI Batch API

Deferred from the deps-update / nested-files / OpenRouter change.