[BFCL] [Bug] the accuracy of multi-turn is lower than bfcl leardboard when using Qwen3 local model inference

[Yayalelelelele]

Describe the issue
When using local Qwen-3(FC) model for inference, the multi-turn acc is lower than bfcl leardboard, while the live and non-live acc is normal.
Same issue as #1145 #1109

local model inference with Qwen3-8b-FC

bfcl leardboard

Reason
When inference, the prompt is not align with the training stage. the code of gorilla/berkeley-function-call-leaderboard/bfcl_eval/model_handler/local_inference/qwen_fc.py lost the content of tool call in the next round

Solutions
update the code of qwen_fc.py

After update, the acc of multi-turn can match with the leardboard