Harden the exercode all-problem check: symlinked problem discovery, close-delay edge, copy-window temp leak, per-answer copy cost, TEST_CASE_RESULT parsing dedup

[exKAZUu]

Deferred findings from the reviews of #318:

1. Symlinked problem directories are silently skipped. visitDirectory in src/cli/checkAllProblems.ts uses entry.isDirectory() / entry.isFile(), which are false for symlinks, so a problem reached only through a symlinked directory (or whose problem.md is a symlink) is skipped without failing the check. Resolve symlinks with stat or fail loudly.
2. close-delay misclassification edge (POSIX). runHarnessProcess settles on the child's close event. If judged code leaves a background process that both escapes the detached process group (e.g. via setsid) and inherits the stdio pipes, a passing run is reported as timed out after 10 minutes even though all TEST_CASE_RESULT lines were printed. Settle on exit plus a short stream-drain grace period, or add a post-kill settle timer.
3. Interrupt during the problem copy leaks the partial temp directory. Temporary roots are registered for signal cleanup only when the harness process spawns; a Ctrl-C while copyProblemDirToTemporaryRoot is still copying leaves the partial problem-utils-isolation_* tree in the OS temp directory. Register the temp root (and install the signal handlers) before the copy starts.
4. The whole problem directory is copied once per model answer. A problem with M answers copies all answers and all test cases M times. Copy once per problem and clone per run (ideally with COPYFILE_FICLONE), and stream-filter TEST_CASE_RESULT lines instead of buffering up to 64 MB of output.
5. TEST_CASE_RESULT line parsing is triplicated. src/cli/checkAllProblems.ts, isAcceptedJudgeOutput in src/helpers/checkProblemDirIsolation.ts, and test/e2e/debugAndJudge.test.ts each parse the lines with slightly different newline handling. Extract a shared parseTestCaseResultLines(stdout) next to src/types/testCaseResult.ts.