[core] allow using initializer allocator for prepack (#26617)

### Description

This PR makes ORT to prefer initializer allocator when calling
`OpKernel::PrePack`.

If an EP does not register an initializer allocator (currently only
WebGPU does this), the behavior is kept unchanged.

### Motivation and Context

Helps to improve the memory usage when doing prepack.

Author: fs-eire

onnxruntime/core/framework/session_state.cc

@@ -588,15 +588,15 @@ Status SessionState::PrepackConstantInitializedTensors(
                   // within this session. Or if the weight is not present on disk,
                   // we store the newly minted pre-packed data.
 
-                  AllocatorPtr session_cpu_alloc = GetAllocator(kernel->Info().GetDevice(OrtMemType::OrtMemTypeDefault));
+                  AllocatorPtr session_initializer_alloc = GetInitializerAllocator(kernel->Info().GetDevice(OrtMemType::OrtMemTypeDefault));
                   PrePackedWeights weights_to_be_filled_in;
                   // The reason we invoke PrePack() before looking into the container for any pre-packed weight
                   // cached by another instance of the same op_type (for the same constant initializer) is because
                   // to truly know if we can use a cached pre-packed weight, we would have to compare the cached
                   // pre-packed weight with the pre-packed weight generated by this instance of the same op_type because
                   // other static properties of the node like node attributes could play a role in the pre-packed
                   // weights' contents.
-                  ORT_RETURN_IF_ERROR(kernel->PrePack(const_initialized_tensor, input_idx, session_cpu_alloc,
+                  ORT_RETURN_IF_ERROR(kernel->PrePack(const_initialized_tensor, input_idx, session_initializer_alloc,
                                                       is_packed,
                                                       &weights_to_be_filled_in));