[Mattral]

docs: simpler quickstart

Removes bandit=BanditConfig(n_arms=1) from the README quickstart and the docs examples wherever one model drafter is
used, and prints which drafter strategy ran. Docs and tests only; no library change. Do not merge until reviewed.

Why it can go

Under the default drafter_strategy="auto", SpeculativeEngine checks config.bandit.n_arms only when you pass a
BanditConfig yourself (check_n_arms = requested != "auto" or "bandit" in config.model_fields_set). With one model drafter,
"auto" resolves to composite-g6-f2, which is one arm, so the default BanditConfig() (n_arms=2) is never compared.

Changes

• README quickstart:
  • drops the BanditConfig import and argument; the config is one line: FlashSpecConfig(device=..., sampling=SamplingConfig(temperature=0.0));
  • adds print("drafter strategy:", engine.drafter_strategy);
  • still 20 lines, ≤ 100 columns.
• README recipes:
  • "Greedy or sampling" and "Turn compile off" drop bandit=BanditConfig(n_arms=1).
  • "Choose a drafter": the lookup-only example keeps it. It sets drafter_strategy="bandit" explicitly, and an explicit
    strategy still checks the arm count against the config. That recipe now imports BanditConfig itself.
• docs/examples.md (the TinyLlama load_drafters example) drops BanditConfig(n_arms=1, strategy="ucb1") and says that
  "auto" runs composite-g6-f2 there.

Proof

New tests/unit/test_readme.py::test_quickstart_auto_runs_composite_and_matches_hf_greedy runs the README quickstart block as
written, on tiny models on CPU. It asserts:

• there is no BanditConfig in the block;
• engine.drafter_strategy == "composite-g6-f2", the reason is "auto: one model drafter…", and the arm is lookup 6 / fallback 2;
• the printed output contains drafter strategy: composite-g6-f2;
• all 128 generated tokens equal the target's generate(do_sample=False) (max_new_tokens=128, no EOS).

The existing README and examples code-block tests still run every block (examples included) with the bandit config removed.

Gate

Local, on this branch:

• make lint (ruff, format, import-linter), mypy --strict flashspec: clean
• pytest tests/unit tests/integration -m "not gpu": 559 passed, 5 skipped
• python scripts/readme_api.py --check: current (no API change)
• CI on this PR: docs-build pass, lint pass, onnx-parity pass, paper pass, test-chaos pass, test-cpu pass

Merging this before #54 or after makes no difference; they touch different files.