Judge tuning (3/5): Preserve native token usage

[ErlisLushtaku]

Keeps the token counts the model reports, so judging cost can be computed from what the judge actually generated, including cached judgements.

• Usage survives inference caching: usage_json is a new column in the completion and judgement caches, restored on cache hits and also written to the meta-eval annotation artifacts together with the new error field.
• Only tokens are cached, not dollar charges. Prices differ per run, so each run applies its own, and a charge stored in an older row is dropped on read.
• Existing databases get the column when they are opened. merge_from keeps the stored usage when an incoming row for the same output has none, so syncing an older cache does not erase counts.

The cost objective in #133 needs this. NePS does not own the cache.

Merge order: #138 → #135 → #140 → #133 → #136.