feat: add miner-bench harness for local miner self-test

[dataCenter430]

Adds miner_bench/: a standalone harness that lets miners evaluate attack strategies against the same scoring rules the mainnet validator applies (faithful port of verify_and_score), without consuming validator resources or polluting on-chain history.

• scoring.py: dependency-light port of PerturbValidator.verify_and_score with the same SSIM/PSNR/Linf/RMSE gates and score combination.
• challenge.py: Pexels fetch + Challenge dataclass; mirrors the validator's challenge generation (prompt rotation, photo sampling, label-from-model).
• strategy.py: MinerStrategy ABC + PGDStrategy reference baseline (textbook untargeted PGD, respects validator's effective max-delta cap).
• runner.py: run_bench() + CLI with per-challenge + aggregate reporting, optional JSON dump for downstream dashboards.
• scripts/bench_miner.py: thin CLI entrypoint.
• README in package: install, CLI flags, library API, plug-in interface, calibration notes vs mainnet.

Smoke-tested end-to-end against live Pexels with EfficientNetV2-M on GPU: 5/5 PGD challenges scored cleanly through the validator-parity scorer.

Test logs:

$ python scripts/bench_miner.py --n-challenges 10 --seed 42 --pgd-steps 25

[bench] loading EfficientNetV2-M on cuda...
[bench] challenge 1/10 ...
  prompt=dog          true=german shepherd           eps=0.0805 attack_ms=  350 score=0.7543 reason=success
[bench] challenge 2/10 ...
  prompt=fish         true=eel                       eps=0.1446 attack_ms=  106 score=0.9151 reason=success
[bench] challenge 3/10 ...
  prompt=cat          true=persian cat               eps=0.1961 attack_ms=  177 score=0.7604 reason=success
[bench] challenge 4/10 ...
  prompt=caprine      true=ibex                      eps=0.1811 attack_ms=  116 score=0.9202 reason=success
[bench] challenge 5/10 ...
  prompt=porcine      true=wild boar                 eps=0.1318 attack_ms=  262 score=0.4492 reason=success
[bench] challenge 6/10 ...
  prompt=amphibian    true=tailed frog               eps=0.1491 attack_ms=  183 score=0.7593 reason=success
[bench] challenge 7/10 ...
  prompt=insect       true=long-horned beetle        eps=0.1225 attack_ms=  131 score=0.9162 reason=success
[bench] challenge 8/10 ...
  prompt=fish         true=goldfish                  eps=0.1441 attack_ms=  128 score=0.9161 reason=success
[bench] challenge 9/10 ...
  prompt=cat          true=egyptian cat              eps=0.1139 attack_ms=  131 score=0.7605 reason=success
[bench] challenge 10/10 ...
  prompt=bird         true=jacamar                   eps=0.0601 attack_ms=   92 score=0.9128 reason=success

strategy        : pgd
challenges      : 10 / requested 10
success rate    : 100.0% (10/10)
mean score      : 0.8064
mean attack_ms  : 168
reasons         :
  success                        10

Sample JSON dump entry (--out results.json)

{
  "task_id": "bench-1779210580533-1",
  "prompt": "dog",
  "true_label": "german shepherd",
  "epsilon": 0.0805,
  "attack_ms": 350,
  "evaluation": {
    "score": 0.7543058425454864,
    "reason": "success",
    "model_prediction": "kelpie",
    "response_time_ms": 350,
    "norm": 0.007843166589736938,
    "rmse": 0.004740638192743063,
    "epsilon": 0.0805,
    "ssim": 0.9930430054664612,
    "psnr_db": 46.4832641328272
  }
}