feat: use pre-generated custom dataset for benchmarking MTP with chat template

[richardhuo-nv]

This is a workaround for adding the chat template for MTP based benchmarking for GLM-5 model.

GLM-5 is using a custom tokenizer, invoking it multiple times will cause pmix failure at runtime.

Workaround is pre-generate the dataset with chat template and use it for MTP based benchmarking.

The config is like this:

benchmark:
  type: "sa-bench"
  isl: 1024
  osl: 1024
  concurrencies: "8192"
  req_rate: "inf"
  dataset_name: "custom"
  dataset_path: "/glm5_datasets/glm5-1024-1024-100000-ratio-1_for_serve.json"
  custom_tokenizer: "glm_moe_dsa"

extra_mount:
  - "/lustre/fsw/core_dlfw_ci/rihuo/glm5_dataset:/glm5_datasets"