mlx-lm issue #1573: RotatingKVCache.to_quantized() raises NotImplementedError, so --kv-bits crashes at first request

mlx_lm.server starts cleanly and /health returns 200, then the first chat completion raises 'RotatingKVCache Quantization NYI' for Gemma 4's hybrid sliding-window layers (35 of 42 layers in 26B-A4B). The guard in maybe_quantize_kv_cache tests hasattr(c, 'to_quantized'), which is satisfied because the method exists and raises, so the failure is deferred to the first request instead of startup. Confirmed on mlx-lm 0.31.3 and current main. PR #1584 (RotatingQuantizedKVCache) was open and mergeable at the time of access; a reporter verified the hybrid Gemma 4 case on an M5 Max with mlx 0.32.0 using that branch.