fix: env and importer reward integrity

[abhishekgahlot2]

Fixes reward-integrity issues in imported and built-in environments.

• BigCodeBench: the runner checks the tests' top-level imports before loading the solution. A module the sandbox lacks now gives no verdict instead of scoring a correct solution 0.
• EvalPlus: a task whose reference fails on any input is not published, and an earlier import's copy is removed. Before, those cases left the denominator, so a solution that failed them could score 1.0.
• file_organizer: each README must hold its heading and description line. Heading-only files no longer earn 1.0.
• cve_2021_44228_log_triage: only the answer's own fields are scored (keywords elsewhere in the output earn nothing). It needs the exact CVE ID and log4shell, counts indicators naming the jndi:ldap and jndi:dns lookups, and gives no remediation credit to a clause that holds a negation.

Each fix has a regression test that fails on main. Affected test files: 110 passed.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}