Add a Vulkan fp16 variant

fp16 is what makes this model worth lowering: at fp32 Vulkan loses to XNNPACK
on an Adreno 840 (0.91x), at fp16 it wins 1.35x, and on a Mali-G76 it is 1.90x.
Halving the file to 218 MB also eases the memory pressure that dominates this
model on 8 GB devices.

Mean pooling is written with keepdim so the reduction stays on the GPU, which
costs one delegate blob less. The 30527-row token embedding is kept off
texture storage, without which fp16 returns silently wrong values
(pytorch/executorch#22333); cosine against the CPU reference is 0.99995 on
Adreno and 0.99997 on Mali.

Requires the ExecuTorch 1.4.1 Vulkan runtime from the v0.10.0-libs-1.4.1
artifacts.

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 all-mpnet-base-v2.pte filter=lfs diff=lfs merge=lfs -text
 all-mpnet-base-v2_xnnpack.pte filter=lfs diff=lfs merge=lfs -text
 xnnpack/all_mpnet_base_v2_xnnpack_fp32.pte filter=lfs diff=lfs merge=lfs -text
+vulkan/all_mpnet_base_v2_vulkan_fp16.pte filter=lfs diff=lfs merge=lfs -text

vulkan/all_mpnet_base_v2_vulkan_fp16.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d891d3f6d3177d751edde4f4c9f436a05f7346addaf865a6ab8431830cf8b1a6
+size 217946754

vulkan/config.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "https://huggingface.co/software-mansion/react-native-executorch-spec/resolve/main/config.schema.json",
+  "model": "all_mpnet_base_v2",
+  "family": "sbert",
+  "capabilities": ["text-embedding"],
+  "backend": "vulkan",
+  "license": "apache-2.0",
+  "variants": [
+    {
+      "file": "all_mpnet_base_v2_vulkan_fp16.pte",
+      "precision": "fp16",
+      "quantized": false,
+      "default": true,
+      "methods": {
+        "forward": {
+          "inputs": [
+            { "shape": [1, 382], "dtype": "int64" },
+            { "shape": [1, 382], "dtype": "int64" }
+          ],
+          "outputs": [{ "shape": [1, 768], "dtype": "float32" }]
+        }
+      }
+    }
+  ]
+}