feat(ai): add Anthropic batch provider

[artell0]

Fixes

Part of #6213 (Incorporate LLMs into Document Ingestion & Processing).

Summary

Adds AnthropicBatchWrapper (cl/ai/llm_providers/anthropic.py), an
Anthropic Message Batches implementation that mirrors the public surface
of GoogleGenAIBatchWrapper so the batch management commands can target
either provider. LLMProvider.ANTHROPIC was already declared and
LLMRequest.api_model_name's help text already references a Claude model —
this fills in the provider itself.

Following the pattern noted in check_gemini_batch_status.py
("If you add other providers (OpenAI, Anthropic), create separate
commands"), I kept this as an independent wrapper alongside google.py.

Two differences from the Gemini path simplify the implementation:

• messages.batches.create accepts the request list directly — there is no
  intermediate JSONL file to upload.
• Prompt caching is implicit prefix caching applied via a cache_control
  breakpoint on the system block — there is no cache object to create or
  look up, so get_or_create_cache has no analog here.

Choices I made (happy to change if you'd prefer otherwise):

• File inputs: inline base64 document blocks (matches the Gemini
  wrapper's input_file_path / input_text task shape). Inline documents
  are capped at 32 MB / 600 pages per request; if your scans routinely
  exceed that, I'll switch to the Files API instead.
• No commands yet: this PR is the provider + tests only, so the shape
  can be reviewed before I add send_anthropic_file_batches /
  check_anthropic_batch_status. The S3/task-staging helpers in the Gemini
  command are provider-agnostic — I can factor them into a shared module
  when wiring the Anthropic command, if that's the direction you want.
• max_tokens defaults to 8192 per task (overridable on
  execute_batch); no default model is hard-coded in the wrapper.

Tests: SimpleTestCase with the Anthropic client mocked (no live calls,
no DB) — model validation, request preparation (inline base64 document +
text, skip-missing-key), system cache_control attachment, status
retrieval, unfinished-batch guard, and result parsing (succeeded + errored).
Mirrors the existing ValidateGeminiModelTest / GoogleGenAIBatchWrapperTest
style.

The Anthropic API key is referenced by env var only and routed through the
same secret mechanism as the Gemini key; nothing is committed.

Deployment

This PR should:

• skip-deploy (skips everything below)
  • skip-web-deploy
  • skip-celery-deploy
  • skip-cronjob-deploy
  • skip-daemon-deploy

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[mlissner]

Hi, I'm not sure this is something we need? Is there more (human) context for this?