feat: migration to mlx-swift-lm v3

MLXServerTests/Server/TokenPrefixCacheQuantizationTests.swift

@@ -249,4 +249,11 @@ private final class NonStandardCache: KVCache {
     ) -> MLXFast.ScaledDotProductAttentionMaskMode {
         .none
     }
+
+    func copy() -> any KVCache {
+        let c = NonStandardCache(tokenCount: 0, headDim: 0)
+        c.state = state
+        c.offset = offset
+        return c
+    }
 }