vLLM Fixes TurboQuant for Qwen 3.5+

Small patch, big practical unblocker: this turns TurboQuant from “broken on modern Qwen hybrids” into something people can actually test in serving setups. The remaining friction is mostly operational, not architectural.

• –The fix targets hybrid-model handling, which is exactly where newer Qwen variants were falling over.
• –Community validation on Qwen 3.6 27B is encouraging, but it is still field testing, not a formal release guarantee.
• –Chunked prefill still needs enough batched tokens to satisfy alignment constraints, so deployers may need to tune `--max-num-batched-tokens`.
• –The new KV cache modes broaden the optimization surface for low-memory inference, especially for large Qwen deployments.
• –This is most relevant to teams already running vLLM in production or benchmarking quantized serving paths.