fix: old version ray placement will set the available GPUs to 0 for vllm

[linqinluli]

1. bug fix for TP>1 in sampling backend. When tensor_parallel_size > 1, vLLM spawns an EngineCore in a background process that inherits CUDA_VISIBLE_DEVICES from the parent Ray actor. Previously, num_gpus=0 caused Ray to set CUDA_VISIBLE_DEVICES="", so the child process saw zero GPUs and vLLM crashed with KeyError: 'bundles'. By assigning num_gpus=tensor_parallel_size, Ray populates CUDA_VISIBLE_DEVICES correctly, and vLLM handles its own internal placement group.
2. Some other modifications for pre-commit and gitignore.