[Feature]: support longer max_num_batched_tokens for lora

[NiuBlibing]

🚀 The feature, motivation and pitch

Support longer max_num_batched_tokens for lora to free long context ablitity of many models have long context such as qwen2(128k).

Alternatives

No response

Additional context

https://github.com/vllm-project/vllm/blob/main/vllm/config.py#L1379-L1386

[jeejeelee]

The current vLLM's LoRA kernel has been replaced with triton-implemented version. You can comment out this snippet and test if it supports up to 128K.

[NiuBlibing]

It works.