Add image benchmarking

[CUHKSZzxy]

Motivation

1. Convenient image benchmarking, adapted from
   https://github.com/sgl-project/sglang/blob/main/python/sglang/bench_serving.py

2. Faster base64 encode / decode
   vLLM & SGLang both adopt pybase64 for faster base64 encode / decode. I would recommend similar updates in lmdeploy.

• https://github.com/vllm-project/vllm/blob/e3a1cd1c59b7cfb8fd6eb05e69393aa7f42dc12d/vllm/multimodal/image.py#L95

• https://github.com/vllm-project/vllm/blob/e3a1cd1c59b7cfb8fd6eb05e69393aa7f42dc12d/requirements/common.txt#L47

• https://github.com/sgl-project/sglang/blob/702426b06ac5675f1c2362b3a8153700c8e05c96/python/sglang/utils.py#L181

Usage

Benchmark usage

python benchmark/profile_restful_api.py \
     --backend lmdeploy-chat \
     --dataset-name image  \
     --num-prompts 100 \
     --random-input-len 100 \
     --random-output-len 100 \
     --random-range-ratio 1 \
     --image-format jpeg \
     --image-count 1 \
     --image-content random \
     --image-resolution 1024x1024 \
     --port 23333