updating vllm modules to have upstream triton attention

[bringlein]

Updating to be able to use vllm-project/vllm#14071.

Vllm and vllm benchmarking runs on A100 and MI250:

# inside docker
VLLM_USE_V1=1 VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 vllm serve /models/llama3.1-8b/instruct/ --disable-log-requests

# new shell inside same container
python3 /scripts/bench_vllm_user_range.py llama3.1-8b/instruct ibm_triton_attn

Please note this only works if the container is build with make dev or make rocm.

[jvlunteren]

Looks good to me.