[Cache Offload] Improve radix cache offload benchmark

[Edenzzzz]

Motivation

1. Improve bench sentence sampling and arg parsing;
2. add README with benchmark instructions
   cc @xiezhq-hermann

Modifications

Checklist

• Format your code according to the Contributor Guide.
• Add unit tests as outlined in the Contributor Guide.
• Update documentation as needed, including docstrings or example tutorials.

[Edenzzzz]

Currently offload seem to decreas the performance on H100.

• context len = 3000
• num groups = 100
• 100 sentences per group
• Command:

python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct \
--host 0.0.0.0 --tensor-parallel-size 4 --enable-hierarchical-cache

Sample from NLTK

• Offload on: 191.94 req/s, hit rate ~79%
• Offload off: 230.87 req/s, hit rate 79.11%

Random Latin text from lorem (original benchmark)

• Offload on: 141.44 req/s, hit rate ~79%
• Offload off: 142.67 req/s, hit rate 79%