Core ML: emit a (1,520,520) int32 index map instead of (1,21,520,520) logits

The argmax now runs inside the graph, built from reduce_max and elementwise ops so it stays on the ANE. 1.05x to 1.63x faster and 7 to 48 percent lower peak memory on an iPhone 16. XNNPACK variants are unchanged.

coreml/config.json

@@ -11,6 +11,8 @@
     {
       "file": "deeplab_v3_mobilenet_v3_large_coreml_fp16.pte",
       "precision": "fp16",
+      "quantized": false,
+      "default": true,
       "size": "mobilenet_v3_large",
       "methods": {
         "forward": {
@@ -29,11 +31,10 @@
             {
               "shape": [
                 1,
-                21,
                 520,
                 520
               ],
-              "dtype": "float32"
+              "dtype": "int32"
             }
           ]
         }
@@ -42,6 +43,8 @@
     {
       "file": "deeplab_v3_resnet50_coreml_fp16.pte",
       "precision": "fp16",
+      "quantized": false,
+      "default": false,
       "size": "resnet50",
       "methods": {
         "forward": {
@@ -60,11 +63,10 @@
             {
               "shape": [
                 1,
-                21,
                 520,
                 520
               ],
-              "dtype": "float32"
+              "dtype": "int32"
             }
           ]
         }
@@ -73,6 +75,8 @@
     {
       "file": "deeplab_v3_resnet101_coreml_fp16.pte",
       "precision": "fp16",
+      "quantized": false,
+      "default": false,
       "size": "resnet101",
       "methods": {
         "forward": {
@@ -91,11 +95,10 @@
             {
               "shape": [
                 1,
-                21,
                 520,
                 520
               ],
-              "dtype": "float32"
+              "dtype": "int32"
             }
           ]
         }