[Code scan] NEB zero failure rates are emitted as missing metrics

This issue was found by a Codex global repository scan of tracked non-test files at commit `8c93925cb10b401b2b83c738bd9263fd74474468`.

### Relevant code
https://github.com/deepmodeling/LAMBench/blob/8c93925cb10b401b2b83c738bd9263fd74474468/lambench/tasks/calculator/neb/neb.py#L62-L86

### Impact
The NEB failure percentage is computed from only the rows where `error > ERROR_THRESHOLD`. If a reaction type has zero failures, that type is absent from the numerator groupby, so the division can produce a missing value instead of an explicit `0.0` failure rate.

A perfect category should be represented as `0.0`, not as a missing metric. Otherwise downstream JSON and leaderboard plots can treat a valid result as absent.

### Suggested fix
Build the denominator counts first, reindex the failure counts against those types with `fill_value=0`, then divide. Add a regression case where one NEB type has no failures.