perf: parallelize independent rules in check --all #81

Context

The engine runs one subprocess per rule, sequentially (engine.evaluate → _run_check). On the normal commit path (a handful of staged files) this is instantaneous and correct. But becwright check --all on a large repo with many rules pays N sequential process starts + N full scans.

(The docs now state this honestly — see the Performance and limits section added in #77 — but stating a limit is not the same as removing it.)

Proposal

Rules are independent by design (each gets its own file list and subprocess), so they can run concurrently:

• concurrent.futures.ThreadPoolExecutor around _run_check calls in evaluate — threads are enough since the work is subprocess-bound, no new dependency needed.
• Keep output ordering deterministic: collect results, print in rule order.
• BECWRIGHT_CHECK_TIMEOUT semantics unchanged (per check).

Acceptance

• check --all wall-clock on a many-rule repo ≈ slowest rule, not the sum
• Output identical to today (same order, same text)
• No new runtime dependency