[Neuron] trim attention kernel tests to fit trn1.2x instance (#14988)

Signed-off-by: Liangfu Chen <liangfc@amazon.com>

tests/neuron/1_core/test_prefix_prefill.py

@@ -314,7 +314,7 @@ def get_active_block_tables(block_tables, query_lens, seq_lens, block_size,
 
         # Test edge cases
         (1, 128, 16, 1024, 4, 2, 16, False),  # large decode batch
-        (16, 4, 8, 8192, 48, 1, 128, True),  # large prefill batch
+        (16, 4, 8, 1024, 4, 2, 128, True),  # large prefill batch
         (4, 12, 32, 2048, 16, 1, 32, True),  # multi-head attention (MHA)
         (4, 12, 32, 2048, 16, 16, 32, True),  # multi-query attention (MQA)
     ])