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

[kinyoklion]

Summary

When data availability is CACHED, every evaluation and every all_flags_state() call logs a warning that the client is using last known values from the feature store. While the client stays in that state, this floods the log with one line per evaluation.

Each of these messages now logs once per client instance. LDClient and AsyncLDClient each hold two flags, one for the evaluation message and one for the all_flags_state() message. The messages say that they are logged once.

In LDClient, a threading.Lock guards the set of each flag. The lock is reached only inside the CACHED branch and only until the flag is set, so the normal evaluation path never touches it. Python has no lock-free atomic flag in the standard library. In AsyncLDClient, a plain boolean is enough because one event loop runs the client and no await separates the check and the set.

The sibling messages for the DEFAULTS case and the availability check itself are unchanged.

---

Note

Overview
Stops log flooding when the SDK serves flags from cached store data before the data source has reached REFRESHED. Warnings for variation / evaluation and for all_flags_state() now emit once per client instead of on every call, with copy noting "This message is logged once."

LDClient adds per-message boolean flags guarded by a threading.Lock (double-checked) so concurrent evaluations only log once. AsyncLDClient uses the same two flags without a lock, relying on single-threaded event-loop semantics between check and set.

Behavior when data is unavailable (non-CACHED) is unchanged. Tests simulate cached-but-uninitialized clients and assert a single warning per message type after repeated calls.

^{Reviewed by Cursor Bugbot for commit 08b2230. Bugbot is set up for automated code reviews on this repo. Configure here.}