Fix integer overflows in attention & cache ops (#1514)

tests/kernels/test_attention.py

@@ -13,7 +13,7 @@ FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
 # This will change depending on the compute capability.
 # - 512 as a buffer
 MAX_SEQ_LEN = get_max_shared_memory_bytes() // FLOAT32_BYTES - 512
-NUM_BLOCKS = 128  # Arbitrary values for testing
+NUM_BLOCKS = 40000  # Arbitrary values for testing
 PARTITION_SIZE = 512
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]