fix: Log the cached-data evaluation warning only once per client

[kinyoklion]

Summary

While data availability is CACHED (the store holds data but the client has not finished initializing), the SDK logs "using last known values" on every evaluation. A client that runs for a long time in that state floods the log with one warning per evaluation.

Each of the two messages (evaluation and all_flags_state) now logs once per LDClient instance. The guard is a Concurrent::AtomicBoolean checked with make_true (a compare-and-set, no lock), the same pattern the context filter uses for its non-symbol attribute warning. The check runs only inside the cached-data branch, so the normal evaluation path is unchanged. Each message now ends with "This message is logged once." to match that precedent. The sibling "store unavailable" messages and the gating conditions are unchanged.

New specs cover: one warning across repeated evaluations, one warning across repeated all_flags_state calls, the two messages counted independently, and a second client logging again.

---

Note

Overview
Stops log flooding when the client evaluates flags or builds all_flags_state while data is CACHED (store has values but initialization has not finished). Those “using last known values” warnings now emit at most once per LDClient, with separate one-shot flags for flag evaluation vs all_flags_state.

Guards use Concurrent::AtomicBoolean#make_true (same idea as other one-time SDK warnings). Wording adds “This message is logged once.”; store-unavailable errors and when evaluations actually use cached data are unchanged.

Adds specs for repeated calls, independent message types, and a fresh warning on a new client instance.

^{Reviewed by Cursor Bugbot for commit f673207. Bugbot is set up for automated code reviews on this repo. Configure here.}