Support Eagle cuda graph for Triton backend

[ispobock]

Motivation

Follow-up of #3466.

# Triton backend Eagle
python3 -m sglang.launch_server --model meta-llama/Llama-2-7b-chat-hf  --speculative-algo EAGLE --speculative-draft lmzheng/sglang-EAGLE-llama2-chat-7B --speculative-num-steps 5 --speculative-eagle-topk 8 --speculative-num-draft-tokens 64 --mem-fraction 0.8 --port 33333 --disable-radix --attention-backend triton

speed: 337.11 token/s