multimodal: length configure should be again .

[kzjeef]

When change model total length by

dashinfer_vlm_serve ... --max_length 64000

it will still limit to 32k by this env:

build/lib/dashinfer_vlm/vl_inference/runtime/qwen_vl.py:
        self.max_input_len = int(getenv("DS_LLM_MAX_IN_TOKENS", "20000"))
        self.max_total_len = int(getenv("DS_LLM_MAX_TOKENS", "32000"))

the length setting should be again, control by command line or env.

eg , qwen vl 2.5 support 128k length in config.json, should read this config to get a better auto settings.

[kzjeef]

Fixed, Thanks!