Enhance vllm support and introduce multi-worker inference

[Carbene]

Features

This PR brings significant upgrades to the vLLM integration and inference performance, along with documentation updates.

1. Full vLLM Support for Train-Free Methods: Expanded the vLLM support range. It now supports all train-free methods, allowing users to leverage vLLM's high-throughput serving for a wider variety of setups. Now it supports caa, sae, sta and vector prompt. Also, merge is fine too. But to ensure the maximal changes, generation part does not support multi-workers, it uses an in-process method to avoid complex hook method.
2. Multi-Worker Inference: Introduced a multi-worker architecture for the application stage. This significantly speeds up the inference process by parallelizing workloads.
3. Documentation Update: Updated readme_2.md to reflect these new capabilities, usage instructions, and configuration changes.

Verification

Successfully running the test in falcon. Observed siginificant result in Qwen3 and Qwen2.5-VL. Test all four methods: CAA, STA, SAE, Vector Prompt.