feat: integrate RL training with vLLM_ascend inference backend

[NengXu001]

Integrates the vLLM inference backend into the RL training pipeline, specifically enabling support for the Ascend NPU environment.

Key Features:

1. Enabled Tensor Parallelism (TP) and Data Parallelism (DP) .
2. Weight Synchronization: Implemented sync_weight logic to ensure consistency between the actor and the inference engine.

Out of Scope:

1. Ref model support
2. Evaluation step
3. Router replay