OpenEnv 0.2.3 conformance: mount /openenv sub-app, add adapter + tests + example client

verifier: fix SIGALRM-in-worker-thread bug that scored every well-formed submission 100/100 under uvicorn (fall back to no-timeout call when signal.signal raises). Trainer was training on a saturated reward landscape; this restores real per-submission scoring.

env: add GET /tasks/{name}/sample_inputs for trainer-side fuzz delegation