Implements LlamaIndex agent runner as requested in openagents-org#264.

## Changes
- Add `LlamaIndexAgentRunner` class in `src/openagents/agents/llamaindex_agent.py`
- Support for LlamaIndex RAG-focused agents (ReActAgent, OpenAIAgent, etc.)
- Tool conversion utilities:
  - `openagents_tool_to_llamaindex()` - convert OpenAgents tools to FunctionTool
  - `llamaindex_tool_to_openagents()` - convert LlamaIndex tools to OpenAgents
- Supports both chat() and query() interfaces
- Async-first with sync fallback
- Event filtering support
- Export new classes from `openagents.agents`

## Usage
```python
from llama_index.core.agent import ReActAgent
from llama_index.llms.openai import OpenAI
from openagents.agents import LlamaIndexAgentRunner

llm = OpenAI(model='gpt-4')
agent = ReActAgent.from_tools(tools, llm=llm)

runner = LlamaIndexAgentRunner(
    llamaindex_agent=agent,
    agent_id='knowledge-agent'
)
runner.start(network_host='localhost', network_port=8600)
```

Closes openagents-org#264

The llamaindex_tool_to_openagents function was using .get() on metadata
assuming it was a dict, but LlamaIndex uses ToolMetadata objects which
have attributes instead. This fix handles both cases properly.