Work out the context limit the way vLLM does #8

The problem

QPF016 and the setup at the end of each report take a model's context limit from max_position_embeddings alone. vLLM works it out in _get_and_verify_max_len and derive_max_model_len_and_key, and three differences matter:

1. It reads other keys too, such as n_positions, max_seq_len, and seq_length, takes the smallest, and lets model_max_length in the model config override them.
2. It takes the smaller of that and model_max_length in tokenizer_config.json.
3. It multiplies by the RoPE scaling factor, except for the su, longrope, llama3, yarn, deepseek_yarn, and deepseek_llama_scaling types, and except for Gemma 3. A linear factor of 4 on a 4,096-token model allows 16,384 tokens, so today QPF016 raises a false alarm at 8,192.

What to do

1. Add a helper in src/quantpreflight/rules.py that returns the limit from the config and the tokenizer config, following the three steps above. The checkpoint's tokenizer config is in meta.tokenizer_config.
2. Use it in _check_qpf016_context_limit and in plan_setup in src/quantpreflight/planner.py.
3. Add a test for each step, including one where the RoPE type is yarn and the factor must not apply.

Test your change

pip install -e ".[dev]"
pytest

New to the project? See CONTRIBUTING.md. Comment here if you get stuck.