The call site of the kernel probably did not check the cu_q_lens[num_seqs[0]] (AKA the actual total num batched q len) is <= max_num_batched_tokens. This could cause hang if we use dynamic grid in kernel.

We should consider adding runtime validation before calling the kernel to prevent cases like this.
For now, we just roll back the dynamic grid change to make the integration work.

Tested:

python test/test_pallas.py -v -k PallasTest.test_ragged_paged_attention_wrapper