Support litellm or langgraph as alternative LLM backends

[benthomasson]

Problem

The LLM integration is currently hardcoded to invoke the claude CLI via subprocess. This works but limits model portability — switching models requires a CLI tool that accepts the same stdin/stdout contract.

Proposed solution

Add support for alternative LLM backends behind the existing MODEL_COMMANDS registry:

• litellm — unified Python API for 100+ LLM providers (OpenAI, Anthropic, Gemini, local models)
• langgraph — for users who want agent-style orchestration

This could be implemented as:

1. A --backend flag (cli, litellm, langgraph) on pipeline/summarize/propose/exam commands
2. Refactor llm.py:invoke_sync to dispatch to the selected backend
3. Keep the CLI subprocess path as the default for backward compatibility

Context

Related to #17 (stringly-typed-llm-integration). The current subprocess-based architecture makes model swapping look trivial but prompt templates may embed implicit model-specific format expectations. A Python SDK backend would enable structured output parsing and better error handling.

Workaround

Currently users can add entries to MODEL_COMMANDS for any CLI tool that accepts prompts on stdin and returns completions on stdout.