`--enforce-eager` option slows down token generation

[mingxin-zheng]

Description

The current script suggests to use --enforce-eager in vllm server

VLM-Surgical-Agent-Framework/scripts/run_vllm_server.sh

Line 16 in 4fa0434

--enforce-eager \

However, this disables the cuda graph acceleration, and as a result, the token generation on A6000 (Holoscan IGX, arm64) is very slow (11.2 tokes/s)

Possible solution

We can add documentation about the flag.

Also, removing the flag will increase the speed 4x (45.9 tokens/s )