Use the new MLX kernels to handle the BF16 matmul. (#2470)

candle-core/src/device.rs

@@ -173,8 +173,8 @@ impl Device {
 
     pub fn supports_bf16(&self) -> bool {
         match self {
-            Self::Cuda(_) => true,
-            Self::Metal(_) | Self::Cpu => false,
+            Self::Cuda(_) | Self::Metal(_) => true,
+            Self::Cpu => false,
         }
     }