feat: LFM2 tool-call normalizer --> fixes #72 (1/10 to 10/10 benchmark)

[Pshyam17]

Problem

LFM2/LFM2.5 emits tool calls in three inconsistent formats depending
on context (Pythonic by default, JSON with a prompt hint, XML in some
Ollama/OpenCode setups). No agentic framework handles all three, forcing
users to write custom proxies.

What this adds

• lfm2_tool_normalizer.py — parses all three formats into a clean
  OpenAI-compatible ToolCall dataclass
• lfm2_agent_loop.py — drop-in agentic loop compatible with plain
  Python callables, LangChain BaseTool, and smolagents Tool
• benchmark.py — before/after comparison across 10 format variants
• quick_start.py — minimal working example

Benchmark

Condition	Success rate
Baseline (naive JSON parse)	1/10
With LFMToolNormalizer	10/10

Notes

Complementary to the home assistant example. This normalizer targets
users running Ollama, OpenCode, or direct HuggingFace inference where
llama.cpp is not in the stack.