feat(gemini): context caching for the page prompt

[jakebromberg]

Summary

• New GeminiClient.create_cache() registers the PAGE_EXTRACTION_PROMPT as a Gemini cachedContent resource (cached as system_instruction).
• extract_page branches on cache presence: when a cache exists, the per-call payload is image-only; the prompt isn't re-sent. When caching is unavailable, the un-cached path is used unchanged.
• core/pipeline.process_pending calls create_cache once at the start of each run; failures degrade silently.
• 6 new client tests (cache lifecycle, idempotency, fallback) + 3 new pipeline tests (cache-once-per-run, cache-failure-continues, no-pending-skips-cache-creation).
• README.md "Cost calibration" section updated to document the caching behavior.

Scope split

The original issue (#40) bundled batch mode + context caching. I'm shipping caching now and proposing batch mode as a follow-up. Reasons:

1. The features are independently valuable and the SDK constraints around each are different. Caching is a per-call payload optimization; batch mode is a 24h-SLA orchestration pattern that doesn't compose with the existing per-page async pipeline (it requires submit → poll → fetch with a new state in jobs.db).
2. The issue's own constraint section says "Don't conflate the two features in tests — they're independent and one should be debuggable without the other."
3. Caching delivers the majority of the Sprint 1 cost win (the prompt is the ~2-3K-token dominant cost; batch is ~50% on top of whatever input rate remains).

A follow-up issue will track batch mode against #37 (Sprint 1 parent). If you'd rather see them combined in one PR, say so and I'll rebase and amend.

SDK reality check

CreateCachedContentConfig has no response_schema field — the schema can only travel in the per-call GenerateContentConfig. So caching saves on the prompt portion only, not on the prompt + schema as the issue body sketched. README and docstrings reflect this.

Partial: #40

Test plan

• All 376 default tests pass (1 deselected).
• ruff check clean.
• ruff format --check clean.
• mypy core cli.py clean.
• Pipeline test confirms caches.create is called exactly once per run with 6 pages.
• Pipeline test confirms the run completes when caches.create raises.
• Pipeline test confirms caches.create is NOT called when there are no pending jobs.
• (Out of scope locally) external_api-marker test against real Gemini caching — needs an opt-in test run with GEMINI_API_KEY.