Rust PyO3 renderer port and native performance paths

[ThomAub]

Summary

Adds the Rust/PyO3 native renderer port with native parity paths.

• NumPy zero copy token APIs
• Renderer sessions and prepared tools

Current benchmark perf

Latest local qwen3 benchmark on this branch:

• native list geomean: 6.25x over Python
• native NumPy geomean: 6.81x over Python
• all-family smoke benchmark completed 327 measured rows

Command used:

uv run python benchmarks/native_vs_python_qwen3.py --families qwen3 --min-time 0.03 --repeats 3 --memory-loops 50

Note

Add Rust/PyO3 native renderer implementations for multiple model families

• Introduces a Rust workspace (crates/) with three members: renderers-core (pure Rust rendering/parsing logic), renderers-py (PyO3 Python bindings compiled as renderers_native), and renderers-cli (a CLI for diffing/benchmarking).
• Adds native renderer implementations for Qwen3, Qwen3.5, Qwen3.6, DeepSeekV3, GLM-4.5/5/5.1, Kimi K2/K2.5, MiniMax M2, Nemotron3, GPT-OSS, and a Jinja-based default renderer, all sharing a common Renderer trait with render, parse_response, stop_token_ids, and bridge_to_next_turn.
• Python renderer classes (Qwen3Renderer, KimiK25Renderer, etc.) now override __new__ to return native Rust instances when RENDERERS_NATIVE env var selects the family and the compiled extension is importable; otherwise they fall back to pure-Python construction.
• Adds _native_router.py for lazy extension loading and tokenizer path resolution, and _native_vision.py for native Qwen-VL image preprocessing with optional NumPy output.
• Adds parity, routing, vision, and NumPy test suites under tests/ and a Rust CI workflow covering fmt, clippy, cargo test, and Miri.
• Risk: __new__ overrides mean instantiating any patched renderer class may return an opaque Rust object; code that introspects instance type or calls Python-only attributes on the returned object will break silently when the native path is active.

^{Macroscope summarized fa47618.}