Forking after HPU driver initialization causes hangs on exit; vLLM auto-overrides to spawn

On HPU, if vLLM's worker start method is set to 'fork' (the upstream default), forking after HPU driver initialization leaves driver state in child processes and can cause hangs on exit. vLLM logs a warning when it auto-overrides to 'spawn'.

The HPU driver (SynapseAI/Habana runtime) is not fork-safe. Forking a process that has already initialized the HPU driver copies the driver state into the child process, which can cause hangs or crashes on exit.

On HPU, vLLM automatically overrides VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. Do not set it to 'fork' explicitly. The 'uni', 'external_launcher', and 'ray' backends do not use Python multiprocessing, so the setting has no effect for them.

This is a Gaudi-specific issue that does not occur on CUDA or ROCm. It affects anyone running multi-card vLLM on Gaudi. The auto-override is handled by the vLLM Gaudi plugin's platform layer.