feat: add Anthropic prompt caching for cost and latency reduction

[ckrough]

Summary

• Enable Anthropic prompt caching via OpenRouter for cost and latency optimization
• Add llm_enable_prompt_caching config option (default: true)
• Log cache hit/write metrics for observability

Changes

• src/config.py: Add llm_enable_prompt_caching setting
• src/infrastructure/llm/openrouter.py: Add cache_control to system messages, log cache metrics
• src/web/routes.py: Pass caching config to provider
• .env.example: Document new config option
• tests/test_llm_provider.py: Add comprehensive tests for caching behavior

Benefits

• Up to 90% reduction in input token costs for cached prompts
• Up to 85% latency reduction for repeated system prompts
• Observability via structured cache metrics logging

Test plan

• All existing tests pass (330 passed)
• Coverage maintained at 80.61%
• mypy strict mode passes
• ruff format/check passes
• bandit security scan passes

Closes: retriever-ghs