Benchmarking : Fix server response aggregation

[varun-sundar-rabindranath]

Summary:
In Streaming mode, the vllm server returns responses as soon as a token is available. However, it doesn't do it in parts, instead, each response is already an aggregate of all the previous responses. Therefore, it is sufficient to record just the last response.

Test:
Manual testing

[andy-neuma]

cool.

[mgoin]

nit: why does this require that the suffix exists and fail if it doesn't? i would simplify it to this if the assertion isn't required

Suggested change

	def trim_suffix(text: str, suffix_str: str) -> str:
	assert len(text) >= len(suffix_str)
	assert text[-1 * len(suffix_str):] == suffix_str
	return text[:-1 * len(suffix_str)]
	def trim_suffix(text: str, suffix_str: str) -> str:
	if text.endswith(suffix_str):
	return text[:-len(suffix_str)]
	return text

[varun-sundar-rabindranath]

It is out of the timing code and I think is good to keep around so we know we aren't trimming any actual data.