fix inference quality caused by temperature parameter in bls

[activezhao]

When the prompt and parameters are the same, I use APIs of ensemble and tensorrt_llm_bls, the results are different.

And the result of ensemble is expected.

I analyzed the code of bls and finally found that the inference quality dropped significantly in some scenarios, because the temperature parameters were not given.

What's more, this problem has led to many bad cases in our prod services.

After fixing the temperature problem, the scores of blue and em are close to vllm of fp16, here is the comparative data:

Here is the code: name_map

And I have added an issue before.
#520