[codex] Add vLLM backend for Qwen LMM adapter

[hansent]

Summary

Adds an env-gated vLLM backend path for the Qwen 3.5 VLM adapter so native Qwen workflow blocks can keep using the existing Inference/LMM request surface while delegating generation to an OpenAI-compatible vLLM server.

Changes

• Adds VLLM_LMM_* environment variables for enabling delegation, configuring the backend URL, served model name, API key, timeout, and temperature.
• Updates InferenceModelsQwen35VLAdapter to skip local AutoModel.from_pretrained(...) when VLLM_LMM_ENABLED=true and call chat.completions.create(...) on the configured vLLM backend instead.
• Converts the native Qwen prompt convention (prompt<system_prompt>system) into OpenAI-compatible chat messages.
• Preserves already-base64 LMM request images by wrapping the existing payload as a data URL, avoiding an extra JPEG recompress before sending to vLLM. Non-base64 image inputs still use the existing image loading fallback.
• Adds unit tests covering URL normalization, prompt splitting, skipped local model loading, vLLM payload construction, and base64 image preservation.

Deploy Notes

For the PlayOn custom image/deployment, the intended env shape is:

VLLM_LMM_ENABLED=true
VLLM_LMM_BASE_URL=http://playon-vllm.playon-vllm.svc.cluster.local:8000
VLLM_LMM_MODEL_NAME=vlm-ocr-14
VLLM_LMM_TEMPERATURE=0
VLLM_LMM_TIMEOUT_SECONDS=120

VLLM_LMM_BASE_URL may omit /v1; the adapter appends it before constructing the OpenAI client.

Validation

uv run --no-project --python /Users/hansent/code/inference/.venv/bin/python pytest \
  tests/inference/unit_tests/models/test_qwen3_5vl_vllm_adapter.py \
  tests/workflows/unit_tests/core_steps/models/foundation/test_qwen_vlm.py \
  tests/workflows/unit_tests/core_steps/models/foundation/test_vlm_remote_execution.py

Result: 36 passed.