fix: make Remote Config LKG writes crash-durable

[shameondev]

Summary

• wait for one atomic SharedPreferences disk commit before reporting Remote Config network success
• preserve the previous committed LKG on commit failure, executor rejection, overflow, or authoritative-removal failure
• keep single-save LRU eviction while making list reconciliation exact and atomic
• fence identity/invalidation races and complete coalesced operations by their semantic postcondition

Reliability evidence

• full sdk:testDebugUnitTest, detektAll, and sdk:assembleDebug pass
• 136 focused manager/cache/durability tests pass
• independent review covered commit failure, coalescing, bounded admission, rejected scheduling, exactly-once callbacks, identity changes, and authoritative removal

Stack

This draft is intentionally based on codex/rc-persistent-lkg / PR #857. It does not add the new Fetched/Active snapshot API; it repairs the durability boundary of the existing whole-context LKG before that cache can be reused.

[shameondev]

Консолидировано в release-train PR (один PR на репозиторий). Ветка сохранена.