…ol search mode

Adds a direct Anthropic Claude API integration as an alternative to OpenAI, with
first-class prompt caching support and a new tool delivery mode.

## Claude Integration
- New `Anthropic` variant in `ModelProvider` enum with base_url and api_key config
- `claude_client.rs`: Direct reqwest-based Anthropic Messages API client supporting
  prompt caching (cache_control breakpoints), streaming (SSE), and native tool use
- `claude_llm.rs`: Full LLM executor implementing the same patterns as the OpenAI
  executor (execute, execute_stream, message mapping, event emission)
- Prompt caching strategy: system prompt, last tool definition, and conversation
  prefix breakpoints for optimal cache hit rates across multi-turn conversations
- Factory function `create_llm_executor()` in llm.rs for provider-based dispatch

## Tool Search Mode
- New `ToolDeliveryMode` enum: `AllTools` (default, send all schemas upfront) vs
  `ToolSearch` (send names+descriptions, agent discovers schemas on demand)
- `tool_search` built-in tool lets agents query for tool schemas by name or keyword
- Added `tool_delivery_mode` to `LlmDefinition` and `StandardDefinition`

https://claude.ai/code/session_01JKpvKsGdQacLQdhSZqX1bK