fix(pt): keep mapping not none during lmp steps when nghost == 0 (#4209)

enhancement on #4144

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

- **New Features**
- Enhanced tensor mapping capabilities with the addition of a new
`mapping_tensor` variable.
- Updated `compute` method to handle ghost atoms and support improved
tensor creation logic.
	- Overloaded `computew` methods to support both double and float types.

- **Bug Fixes**
- Improved error handling in the `translate_error` method for better
exception management.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

source/api_cc/include/DeepPotPT.h

@@ -338,6 +338,7 @@ class DeepPotPT : public DeepPotBase {
   int do_message_passing;  // 1:dpa2 model 0:others
   bool gpu_enabled;
   at::Tensor firstneigh_tensor;
+  c10::optional<torch::Tensor> mapping_tensor;
   torch::Dict<std::string, torch::Tensor> comm_dict;
   /**
    * @brief Translate PyTorch exceptions to the DeePMD-kit exception.

source/api_cc/src/DeepPotPT.cc

@@ -164,7 +164,6 @@ void DeepPotPT::compute(ENERGYVTYPE& ener,
   std::vector<std::int64_t> atype_64(datype.begin(), datype.end());
   at::Tensor atype_Tensor =
       torch::from_blob(atype_64.data(), {1, nall_real}, int_option).to(device);
-  c10::optional<torch::Tensor> mapping_tensor;
   if (ago == 0) {
     nlist_data.copy_from_nlist(lmp_list);
     nlist_data.shuffle_exclude_empty(fwd_map);