Filter empty prompt in random bench serving

[ispobock]

Motivation

Fix issue:

 RuntimeWarning: divide by zero encountered in scalar floor_divide
  ratio = (input_lens[i] + prompt_len - 1) // prompt_len

Reproduce:

python3 -m sglang.bench_serving --backend sglang --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json --dataset-name random --random-input 128 --random-output 64 --num-prompts 3200 --request-rate 32 --random-range-ratio 1.0

[ispobock]

By the way, if an empty prompt request is sent while other requests are in decoding, the server will fail for both normal and overlap cases:

[2024-11-12 11:34:43 TP0] Traceback (most recent call last):
  File "/workdir/repos/sglang/python/sglang/srt/managers/scheduler.py", line 1210, in run_scheduler_process
    scheduler.event_loop_overlap()
  File "/workdir/tools/miniconda/miniconda3/envs/sgl-vl/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/workdir/repos/sglang/python/sglang/srt/managers/scheduler.py", line 368, in event_loop_overlap
    batch = self.get_next_batch_to_run()
  File "/workdir/repos/sglang/python/sglang/srt/managers/scheduler.py", line 615, in get_next_batch_to_run
    self.running_batch.merge_batch(self.last_batch)
  File "/workdir/repos/sglang/python/sglang/srt/managers/schedule_batch.py", line 959, in merge_batch
    raise e
  File "/workdir/repos/sglang/python/sglang/srt/managers/schedule_batch.py", line 956, in merge_batch
    self.output_ids = torch.concat([self.output_ids, other.output_ids])
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument tensors in method wrapper_CUDA_cat)

cc: @merrymercy