Core ML: replace the int8 variant with fp16

The int8 build applied weight-only quantization on top of compute_precision=FLOAT32. That combination took the accuracy cost of quantization and none of the benefit: fp32 compute is the one precision the Apple Neural Engine refuses, so the model was locked onto CPU/GPU.

Measured on an iPhone 16, ExecuTorch 1.4.1, Release, 3 warmup plus 15 timed runs, median:
fp16 is 1.6x (50.3 ms -> 31.0 ms) faster than int8.

int8 was also measurably less accurate: its confident detections (score > 0.3) agreed with fp16 on only 4.3% of confident boxes at IoU 0.9. fp16 is the higher precision build, so the divergence is int8 degrading rather than fp16.

Cost: file size roughly doubles (31.6 -> 62.3 MB). int8 is removed rather than kept as a smaller option, because slower and visibly different is not a trade worth offering.

v0.9.0 and below are untouched.

coreml/config.json

@@ -1,18 +1,17 @@
 {
   "$schema": "https://huggingface.co/software-mansion/react-native-executorch-spec/resolve/main/config.schema.json",
-  "model": "rfdetr_nano",
+  "model": "rfdetr_nano_segmentation",
   "family": "rfdetr",
   "capabilities": [
-    "object-detection",
     "instance-segmentation"
   ],
   "backend": "coreml",
   "license": "apache-2.0",
   "variants": [
     {
-      "file": "rfdetr_nano_coreml_int8.pte",
-      "precision": "int8",
-      "quantized": true,
+      "file": "rfdetr_nano_coreml_fp16.pte",
+      "precision": "fp16",
+      "quantized": false,
       "default": true,
       "methods": {
         "forward": {
@@ -30,7 +29,6 @@
           "outputs": [
             {
               "shape": [
-                1,
                 100,
                 4
               ],
@@ -38,15 +36,18 @@
             },
             {
               "shape": [
-                1,
-                100,
-                2
+                100
+              ],
+              "dtype": "float32"
+            },
+            {
+              "shape": [
+                100
               ],
               "dtype": "float32"
             },
             {
               "shape": [
-                1,
                 100,
                 78,
                 78

coreml/{rfdetr_nano_coreml_int8.pte → rfdetr_nano_coreml_fp16.pte}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a6d129477e4f16ea5a1126aef82a139a03997a143c2a4edaf514e1e215d0cb05
-size 31560333
+oid sha256:5882a3124832a8e09c63c5395d0e855a9d7acb23c56309efdddc2f1a8ed3d62c
+size 62341694