Add multi-model benchmark: Gemini 2.5 Pro, 3 Flash, 3.5 Flash

[KhanMih]

Multi-model benchmark addressing single-model evaluation feedback.

Summary

Addresses reviewer feedback that the study is a single-model evaluation. This PR runs the 30-task benchmark across three Gemini models and presents side-by-side results.

Changes

• Bug fix — Fixed `PlanEngine` to handle `langchain-google-genai` v4+ response format where `response.content` returns a list of content blocks instead of a plain string. Without this fix, Gemini 3.x models fail on every task with JSON parse errors.
• New script — `examples/compare_models.py` loads multiple benchmark result JSONs and produces Rich terminal tables, McNemar cross-model tests, and markdown-ready output for the README.
• README update — Key Results section now shows a side-by-side three-model comparison table with real benchmark numbers. Original paper results preserved in a collapsible block.
• CLI update — `_print_paper_results_table()` now displays multi-model Dynamic Retrieval results.
• Makefile — Added `benchmark-multi` target to run all three models sequentially.

Files changed

• `src/behavioral_memory/planner/engine.py` — `_extract_text()` helper for v4+ content blocks
• `src/behavioral_memory/cli.py` — multi-model results table
• `examples/compare_models.py` — new comparison script
• `README.md` — multi-model results + reproduction instructions
• `Makefile` — `benchmark-multi` target