M21-P1: Anthropic prompt caching for ClaudeProvider

[bug-ops]

Parent: #336

Summary

Enable Anthropic prompt caching by sending structured system content blocks with cache_control markers. Expected to reduce input tokens by 80-90%.

Requirements

1. Add anthropic-beta: prompt-caching-2024-07-31 header to all Claude API requests
2. Convert system field from Option<&str> to Option<Vec<SystemContentBlock>> in request bodies
3. Split system prompt into cacheable blocks:
   • Block 1 (cached): base prompt + active skills
   • Block 2 (cached): tool catalog + environment context
   • Block 3 (not cached): project configs, repo map, MCP prompt
4. Inject section markers in rebuild_system_prompt for splitting
5. Parse usage.cache_read_input_tokens from responses

Acceptance Criteria

• All Claude requests include anthropic-beta header
• System prompt sent as structured content blocks with cache_control
• Cache hit rate visible in metrics/logs
• Existing tests pass, no trait-level breaking changes

Files

• crates/zeph-llm/src/claude.rs
• crates/zeph-core/src/agent/context.rs