fix: refresh stale repo weights on cached evaluations (#364)

[jonathanchang31]

Summary

• Cached miner evaluations retained repo_weight_multiplier from the round they were originally scored. When master_repositories.json was updated between rounds, cached evaluations used stale weights for scoring, causing incorrect repository weights.
• Added _refresh_cached_repo_weights() in reward.py that updates repo_weight_multiplier on all PRs in cached evaluations using the current master_repositories before finalize_miner_scores() runs.
• DB storage still skips cached UIDs (to avoid overwriting file patches with NULL from lightweight cache copies) and self-corrects on the next successful evaluation.

Fixes #364

Test plan

• 10 new unit tests covering: merged/open/closed PRs, multi-repo scenarios, rounding, fallback default for removed repos, non-cached UIDs untouched, mixed cached+fresh evaluations
• Full test suite (282 tests) — all pass, no regressions
• Ruff lint — all checks passed
• Ruff format — all files formatted
• Pyright type check — 0 errors, 0 warnings

[anderdc]

if master repositories is updated mid round, it's from a push from test -> main, which causes the validator to restart and thus all cached evaluations are cleared/reset.

The only case this happens is if a validator is trying to make their own repository weights, which would then cause them to lose vtrust, so this scenario you're fixing doesn't happen in production.