[GPT-OSS] Accuracy and performance config differences - Reasoning effort

[attafosu]

Reference in the regular inference uses "low" and "high" reasoning efforts for performance and accuracy runs respectively. The max tokens are also 10k and 32k respectively - reported in #303
The current implementations for the gpt-oss dataset presets and also the message specs (both sglang and vllm) do not account for these.
If we want parity/comparability with the legacy inference submissions we'd need to find a way to resolve this.

@nvzhihanj @arekay-nv

[attafosu]

After some offline discussions with @tianmu-li we may be able to resolve these issues by wiring "reasoning_effort" parameter to dataset config and amending (adding it to) the openai chat completions adapter accordingly. The gptoss presets would also see some refactoring as well.
The issue with different max tokens for perf and acc would also be solved by moving the "max_new_tokens" to dataset config.

However, such changes to the configs/schema may be less desirable this late towards submission

[nvzhihanj]

Thanks, yes I think using the effort and max_OSL hint in the API might be the simplest way around.

cc: @arekay-nv to check if this is already supported