Don't recompute numel and contiguous in detach (#112689)

When symolic shapes are involved, `refresh_numel` and `refresh_contiguous` are
fairly expensive since they dispatch to python for SymInt handling. However, in
the case of detach we can just copy the existing values instead.
Pull Request resolved: #112689
Approved by: https://github.com/lezcano, https://github.com/ezyang

c10/core/TensorImpl.cpp

@@ -819,9 +819,6 @@ c10::intrusive_ptr<TensorImpl> TensorImpl::shallow_copy_and_detach_core(
       /*dest_impl=*/impl.get(),
       /*version_counter=*/std::forward<VariableVersion>(version_counter),
       /*allow_tensor_metadata_change=*/allow_tensor_metadata_change);
-
-  impl->refresh_numel();
-  impl->refresh_contiguous();
   return impl;
 }
 
@@ -871,6 +868,7 @@ void TensorImpl::copy_generic_tensor_metadata(
       src_impl->is_non_overlapping_and_dense_;
   dest_impl->is_wrapped_number_ = src_impl->is_wrapped_number_;
   dest_impl->reserved_ = src_impl->reserved_;
+  dest_impl->numel_ = src_impl->numel_;
   if (src_impl->extra_meta_ != nullptr) {
     dest_impl->extra_meta_ = src_impl->extra_meta_->clone();
   }

c10/core/TensorImpl.h

@@ -2071,8 +2071,6 @@ struct C10_API TensorImpl : public c10::intrusive_ptr_target {
         /*dest_impl=*/this,
         /*version_counter=*/version_counter(),
         /*allow_tensor_metadata_change=*/allow_tensor_metadata_change());
-    refresh_numel();
-    refresh_contiguous();
   }
 
   // Inference tensor doesn't have version counter,