fix: use --runner pooling instead of --task embed for vllm serve

Fix vllm serve command for different versions

Add vLLM server mode: start once, route from anywhere

vLLM-based route_text(), uv install guide, complete runnable examples

Fix: use transformers AutoModel instead of sentence_transformers for Qwen3 compatibility

Fix Quick Start: complete runnable example with embedding generation

Fix Quick Start: use snapshot_download for proper import

Initial release: R2-Router (#1 on RouterArena)

initial commit