feat: add vllm_python config to specify vLLM interpreter
Allow VLLMRollout to use a different Python interpreter for the vLLM
subprocess via the vllm_python field in RolloutConfig. Useful when
vLLM is installed in a separate virtual-env or conda environment.
Also fix stderr handling in _wait_for_health: read full stderr on
crash and show the tail; redirect stdout to DEVNULL to avoid PIPE
buffer deadlock.
Signed-off-by: Guokai Ma <guokai.ma@intel.com>
