cuda graph + num-scheduler-steps bug fix

[seungrokj]

This is the same approach (vllm-project#8340) in vllm/attention/backends/utils.py to prevent the following error:

[rank0]:   File "/opt/conda/envs/py_3.9/lib/python3.9/site-packages/vllm/worker/model_runner.py", line 869, in build
[rank0]:     attn_metadata = self.attn_metadata_builder.build(
[rank0]:   File "/opt/conda/envs/py_3.9/lib/python3.9/site-packages/vllm/attention/backends/utils.py", line 223, in build
[rank0]:     input_block_tables[i, :len(block_table)] = block_table
[rank0]: ValueError: could not broadcast input array from shape (257,) into shape (256,)

@shajrawi @JArnoldAMD @andyluo7

[shajrawi]

thanks for the fix

[shajrawi]

Can you please the linter so we can merge it?