Bump SessionConcurrencyTest ctx size to 4096

The previous ctxSize=512 was too tight for the concurrent-send test:
4 threads x 3 calls accumulate into a single Session transcript and
each send() resubmits the full history, so by ~call 10 the prompt
exceeds 512 tokens and the native side throws LlamaException. Bumping
to 4096 (well under the 16384 trained ctx) keeps the whole 12-turn
transcript well within bounds.

Author: claude

src/test/java/net/ladenthin/llama/SessionConcurrencyTest.java

@@ -57,7 +57,7 @@ public static void setup() {
         int gpuLayers = Integer.getInteger(TestConstants.PROP_TEST_NGL, TestConstants.DEFAULT_TEST_NGL);
         model = new LlamaModel(
                 new ModelParameters()
-                        .setCtxSize(512)
+                        .setCtxSize(4096)
                         .setModel(TestConstants.MODEL_PATH)
                         .setGpuLayers(gpuLayers)
                         .setFit(false)