M1: Agent loop with the native tool protocol

The core loop: send messages, receive a response, dispatch any tool calls, append results, repeat until the model answers or a step limit is hit. Native OpenAI-style tool_calls only — the text protocol is issue for M2.

Guard rails the loop needs from the start, because small models hit all of them: a hard step ceiling, detection of a model repeating an identical tool call, and a clear error when the model emits malformed tool arguments rather than a stack trace.

Done when liza --provider lmstudio "read hello.py and tell me what it does" completes a request -> tool call -> tool result -> answer cycle against a local model.

Start at the liza CLI path used by liza --provider lmstudio and trace how provider responses and native tool_calls are handled. Implement the request-to-tool-result-to-answer cycle with a step ceiling, repeated-call detection, and clear malformed-argument errors; done is the documented read hello.py example completing against a local model.