responses api in vllm worker async

[cmunley1]

Is your feature request related to a problem? Please describe.

add responses api endpoint in nemo_rl/models/generation/vllm/vllm_worker_async.py for training models natively on responses api like gpt-oss with nemo gym environments

[cmunley1]

i added this locally, but run into some issues. for non gpt-oss model:

  ESC[36m(VllmAsyncGenerationWorker pid=231633, ip=10.66.4.76)ESC[0m [/v1/responses] ErrorResponse: 
   {'error': {'message': 'enable_response_messages is currently only supported for gpt-oss',        
  'type': 'BadRequestError', 'param': None, 'code': 400}}   

for gpt-oss:

  ESC[36m(VllmAsyncGenerationWorker pid=200641, ip=10.66.4.35)ESC[0m [/v1/responses] ErrorResponse: 
   {'error': {'message': 'logprobs are not supported with gpt-oss models', 'type':             

the first i believe is resolved in newer versions of vllm (vllm-project/vllm#29549). i am using nightly container. will revisit when we update vllm.