[ExecuTorch][WebGPU] Graph-owned scratch buffers for fused-op intermediates

Pull Request resolved: pytorch#20073

Add `WebGPUGraph::create_scratch_buffer` for fused-op intermediates (SDPA's `attn_weights`/`attn_weights_softmax`) that are not model tensors and live only between dispatches. Graph-owned, released in the destructor. Vulkan models these as graph tensors; we use raw buffers (buffer-only backend). Consumed by the fused SDPA op above.
ghstack-source-id: 391378805
@exported-using-ghexport

Differential Revision: [D107543259](https://our.internmc.facebook.com/intern/diff/D107543259/)

Author: JulianCloudNTH

backends/webgpu/runtime/WebGPUGraph.cpp

@@ -48,6 +48,17 @@ size_t vk_datatype_size(vkgraph::VkDataType dtype) {
 
 WebGPUGraph::WebGPUGraph() = default;
 
+WGPUBuffer WebGPUGraph::create_scratch_buffer(size_t nbytes) {
+  WGPUBufferDescriptor buf_desc = {};
+  buf_desc.size = nbytes > 0 ? nbytes : 4;
+  buf_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst |
+      WGPUBufferUsage_CopySrc;
+  buf_desc.mappedAtCreation = false;
+  WGPUBuffer buffer = wgpuDeviceCreateBuffer(device_, &buf_desc);
+  scratch_buffers_.push_back(buffer);
+  return buffer;
+}
+
 WebGPUGraph::~WebGPUGraph() {
   for (size_t i = 0; i < tensors_.size(); i++) {
     if (tensors_[i].buffer &&
@@ -60,6 +71,11 @@ WebGPUGraph::~WebGPUGraph() {
       wgpuBufferRelease(buf);
     }
   }
+  for (auto& buf : scratch_buffers_) {
+    if (buf) {
+      wgpuBufferRelease(buf);
+    }
+  }
   for (auto& buf : output_staging_buffers_) {
     if (buf) {
       wgpuBufferRelease(buf);

backends/webgpu/runtime/WebGPUGraph.h

@@ -119,6 +119,9 @@ class WebGPUGraph {
     uniform_buffer_bytes_ += bytes;
   }
 
+  // Graph-owned scratch storage buffer for fused-op intermediates (e.g. SDPA).
+  WGPUBuffer create_scratch_buffer(size_t nbytes);
+
   WGPUShaderModule get_or_create_shader(
       const std::string& key,
       const char* wgsl_source);
@@ -173,6 +176,9 @@ class WebGPUGraph {
   std::vector<WGPUBuffer> shared_buffers_;
   std::vector<size_t> shared_buffer_sizes_;
 
+  // Long-lived scratch storage buffers for fused ops (e.g. SDPA temporaries).
+  std::vector<WGPUBuffer> scratch_buffers_;
+
   // Staging buffers for reading back outputs (MapRead | CopyDst).
   std::vector<WGPUBuffer> output_staging_buffers_;