Replicate network weights only to used NUMA nodes

On a system with multiple NUMA nodes, this patch avoids unneeded replicated
(e.g. 8x for a single threaded run), reducting memory use in that case.

Lazy initialization forced before search.

Passed STC:
https://tests.stockfishchess.org/tests/view/66a28c524ff211be9d4ecdd4
LLR: 2.96 (-2.94,2.94) <-1.75,0.25>
Total: 691776 W: 179429 L: 179927 D: 332420
Ptnml(0-2): 2573, 79370, 182547, 78778, 2620

closes https://github.com/official-stockfish/Stockfish/pull/5515

No functional change

src/search.cpp

@@ -127,6 +127,12 @@ Search::Worker::Worker(SharedState&                    sharedState,
     clear();
 }
 
+void Search::Worker::ensure_network_replicated() {
+    // Access once to force lazy initialization.
+    // We do this because we want to avoid initialization during search.
+    (void) (networks[numaAccessToken]);
+}
+
 void Search::Worker::start_searching() {
 
     // Non-main threads go directly to iterative_deepening()