feat(client.py): data preparation and agent impl

[agronskiy]

• can be run now with actual agent impl

• adds two modes to the client.py

  • data preparation (downloads and runs)
  • running (for dev/debug purposes)

• because I pipecleaned on integrate.api.nvidia.com, some things turned out related to some servers rejecting payloads with Nones, corrected in vllm translator layer

 ng_run "+config_paths=[resources_servers/bash_sandbox/configs/bash_sandbox.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]" \
      "policy_base_url=https://integrate.api.nvidia.com/v1" \
      "policy_api_key=${NVIDIA_API_KEY}" \
      "policy_model_name=nvdev/nvidia/llama-3.3-nemotron-super-49b-v1"

# this in another terminal
python responses_api_agents/gdpval_agent/client.py prepare \
      --output-jsonl resources_servers/bash_sandbox/data/train.jsonl \
      --split train \
      --limit 2

ng_collect_rollouts \
      +agent_name=bash_sandbox_agent \
      +input_jsonl_fpath=resources_servers/bash_sandbox/data/train.jsonl \
      +output_jsonl_fpath=output/gdpval_rollouts.jsonl

[vadam5]

lgtm