feat: add vllm rust frontend

[gufengc]

Summary

Adds vLLM Rust frontend support as Parallax's HTTP entrypoint and wires executors to the vLLM engine-core MessagePack/ZMQ protocol.

What changed

• Launches vllm-rs frontend from parallax.launch and adds a build helper for installing the Rust frontend binary.
• Adds engine-core protocol helpers for request decoding, sampling parameter mapping, routing metadata extraction, ready registration, token outputs, terminal outputs, aborts, and errors.
• Updates MLX, SGLang, and vLLM executors to emit vLLM engine-core outputs instead of the old Python HTTP-server payloads.
• Forwards Parallax routing metadata through vllm_xargs and updates frontend streaming handling for reasoning deltas.
• Refreshes frontend dist assets and replaces old HTTP handler tests with backend request and engine-core protocol coverage.

Impact

Parallax now relies on the official vLLM Rust frontend for OpenAI-compatible HTTP serving. Backend requests carry scheduler/routing metadata through supported vLLM extension arguments, while executor outputs use vLLM-compatible finish reasons and finished-request lists.

Validation

• uv run --no-sync --with pre-commit pre-commit run --all-files
• uv run --no-sync --with pytest --with pytest-mock --with pytest-cov pytest