Use temp allocator for kernel registry (pytorch#13012)

Summary:
Pull Request resolved: pytorch#13012

When indexing to the registry to get the op, memory is allocated from the method allocator to instantiate some TensorMeta and included objects. This memory is only used for that purpose and is not needed for the entire lifetime of the Method. Thus, we can instead use temp allocator which can later be reset and free up memory as needed.

Differential Revision: D79285675

Author: JakeStevens

runtime/executor/method.cpp

@@ -670,7 +670,7 @@ Error Method::resolve_operator(
     size_t kernel_index,
     InstructionArgs args,
     size_t n_args) {
-  // TODO(T153505381, T153506819) Investigate optimizing this function for both
+  // TODO(T153506819) Investigate optimizing this function for both
   // space and time.
 
   // resolve name
@@ -691,8 +691,11 @@ Error Method::resolve_operator(
   }
 
   // resolve tensor meta
-  auto method_allocator = memory_manager_->method_allocator();
-  TensorMeta* meta = method_allocator->allocateList<TensorMeta>(n_args);
+  auto allocator = memory_manager_->temp_allocator();
+  if (allocator->size() == 0) {
+    allocator = memory_manager_->method_allocator();
+  }
+  TensorMeta* meta = allocator->allocateList<TensorMeta>(n_args);
   if (meta == nullptr) {
     return Error::MemoryAllocationFailed;
   }
@@ -705,8 +708,7 @@ Error Method::resolve_operator(
       auto tensor = eval->toTensor();
       meta[count].dtype_ = tensor.scalar_type();
       executorch::aten::DimOrderType* dim_order_ptr =
-          method_allocator->allocateList<executorch::aten::DimOrderType>(
-              tensor.dim());
+          allocator->allocateList<executorch::aten::DimOrderType>(tensor.dim());
       if (dim_order_ptr == nullptr) {
         return Error::MemoryAllocationFailed;
       }