[webgpu] Use LazyRelease for prepack allocator

onnxruntime/core/providers/webgpu/webgpu_context.cc

@@ -138,10 +138,10 @@ void WebGpuContext::Initialize(const WebGpuContextConfig& config) {
                                                config.buffer_cache_config.uniform.mode,
                                                config.buffer_cache_config.query_resolve.mode);
 
-    // create initializer buffer manager. cache is always disabled for initializer buffer manager
+    // create initializer buffer manager.
     initializer_buffer_mgr_ = BufferManagerFactory::Create(*this,
-                                                           BufferCacheMode::Disabled,
-                                                           BufferCacheMode::Disabled,
+                                                           BufferCacheMode::LazyRelease,
+                                                           BufferCacheMode::LazyRelease,
                                                            BufferCacheMode::Disabled);
 
     // create program manager

onnxruntime/core/providers/webgpu/webgpu_kernel.cc

@@ -49,6 +49,12 @@ Status WebGpuKernel::PrePack(const Tensor& tensor, int input_idx, AllocatorPtr /
 
   Status s = PrePackInternal(context, tensor, input_idx, ep_.PrepackAllocator(), is_packed);
 
+  if (is_packed) {
+    // Flush pending commands to ensure GPU buffer creations are completed.
+    // This allows the initializer buffer manager to release temporary buffers and reduce memory usage.
+    webgpu_context_.Flush(webgpu_context_.InitializerBufferManager());
+  }
+
   if (webgpu_context_.ValidationMode() >= ValidationMode::Full) {
     ORT_RETURN_IF_ERROR(webgpu_context_.PopErrorScope());
   }