Regenerate FROZEN_EVAL_SCORED.jsonl with shipped V43

The score and tier fields were precomputed from an older model in the
v0.0.1 release. Re-scored every row with the currently shipped ONNX
(models/l5e/model.int8.onnx) so scripts/bench_oss.py compares the
correct numbers.

On the same eval slice (791 attacks / 132 benigns):
  recall@0.95  83.94%  (was 33.63% with stale scores)
  FPR@0.95     10.61%  (was 25.00%)

Author: sandeepl337