fix: make nCoinCacheUsage std::atomic to close an adaptive-dbcache data race

The scheduled AdjustCoinCacheForMemoryPressure task writes nCoinCacheUsage from
the scheduler thread holding no lock, while cs_main-holding threads
(FlushStateToDisk, VerifyDB) read it -- an unsynchronized read/write of a
non-atomic size_t (C++ UB). Make it std::atomic<size_t>; correct the comment
that incorrectly described the access as lock-free/race-free.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/main.cpp

@@ -109,7 +109,7 @@ bool fIsBareMultisigStd = true;
 bool fCheckBlockIndex = false;
 bool fCheckpointsEnabled = true;
 bool fCoinbaseEnforcedProtectionEnabled = true;
-size_t nCoinCacheUsage = 5000 * 300;
+std::atomic<size_t> nCoinCacheUsage(5000 * 300);
 uint64_t nPruneTarget = 0;
 // If the tip is older than this (in seconds), the node is considered to be in initial block download.
 int64_t nMaxTipAge = DEFAULT_MAX_TIP_AGE;