[Bugfix][TPU] Use np array when updating cache slot_mapping (vllm-project#17971)

Signed-off-by: Siyuan Liu <lsiyuan@google.com>
Signed-off-by: minpeter <kali2005611@gmail.com>

Author: lsy323

vllm/v1/worker/tpu_model_runner.py

@@ -531,7 +531,7 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"):
         np.add(block_numbers * self.block_size,
                block_offsets,
                out=self.input_batch.block_table.
-               slot_mapping_cpu[:total_num_scheduled_tokens])
+               slot_mapping_np[:total_num_scheduled_tokens])
 
         # Prepare the attention metadata.
         self.query_start_loc_np[0] = 0