fix(pt_expt): let deepmd.pt import errors propagate in comm op check (#5474)

## Summary

`_check_underlying_ops_loaded()` in `deepmd/pt_expt/utils/comm.py` wraps
`import deepmd.pt` in a blanket `except Exception: pass`, then falls
through to a generic `RuntimeError` telling the user to build
`libdeepmd_op_pt.so`.

The problem: when the .so *is* built but loaded against a mismatched
torch version, `import deepmd.pt` raises an `ImportError` with
diagnostic detail (e.g. `undefined symbol: ...`) — exactly the message
the user needs. The current code hides it and tells them to rebuild a
library that's already built.

This PR removes the `try/except` and lets the import error propagate.
The downstream `RuntimeError` still fires for the case where the import
succeeds but the ops still aren't registered.

## Trade-off

External callers that previously caught `RuntimeError` from `comm.py`
import will now see the raw `ImportError` for the .so-mismatch case. No
in-tree caller does this. The diagnostic gain outweighs the contract
change.

## Test plan

- [x] Existing pt_expt tests (every consumer imports `comm.py`) — happy
path unchanged
- [ ] CI green

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

* **Bug Fixes**
* Improved initialization error reporting: when native registrations
fail to load, the underlying import/ABI error is now preserved and
surfaced instead of being masked by a generic message, making root
causes clearer for troubleshooting.

<!-- review_stack_entry_start -->

[![Review Change
Stack](https://storage.googleapis.com/coderabbit_public_assets/review-stack-in-coderabbit-ui.svg)](https://app.coderabbit.ai/change-stack/deepmodeling/deepmd-kit/pull/5474?utm_source=github_walkthrough&utm_medium=github&utm_campaign=change_stack)

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

---------

Co-authored-by: Han Wang <wang_han@iapcm.ac.cn>

Author: wanghan-iapcm

deepmd/pt_expt/utils/comm.py

@@ -50,25 +50,38 @@ def _check_underlying_ops_loaded() -> None:
     like DDP-spawned subprocesses that re-import modules from scratch
     and never see the test conftest's ``import deepmd.pt``.
     """
-    if not (
-        hasattr(torch.ops, "deepmd_export")
-        and hasattr(torch.ops.deepmd_export, "border_op")
-        and hasattr(torch.ops.deepmd_export, "border_op_backward")
-    ):
+
+    def _ops_registered() -> bool:
+        return (
+            hasattr(torch.ops, "deepmd_export")
+            and hasattr(torch.ops.deepmd_export, "border_op")
+            and hasattr(torch.ops.deepmd_export, "border_op_backward")
+        )
+
+    import_err: Exception | None = None
+    if not _ops_registered():
         # Triggers cxx_op.py which torch.ops.load_library's the .so.
         try:
             import deepmd.pt  # noqa: F401
-        except Exception:
-            # If deepmd.pt itself fails to import, fall through to the
-            # explicit RuntimeError below — clearer than re-raising a
-            # potentially-unrelated import error.
-            pass
-
-    if not (
-        hasattr(torch.ops, "deepmd_export")
-        and hasattr(torch.ops.deepmd_export, "border_op")
-        and hasattr(torch.ops.deepmd_export, "border_op_backward")
-    ):
+        except Exception as exc:
+            # ``deepmd/pt/__init__.py`` loads ``cxx_op`` (which registers
+            # the ops) before running ``load_entry_point("deepmd.pt")``.
+            # A broken third-party entry point can make the import raise
+            # *after* the ops were already registered, so only re-raise
+            # when the registration is still missing — that branch is the
+            # one where the error (typically an ``undefined symbol`` ABI
+            # mismatch against libdeepmd_op_pt.so) carries the diagnostic
+            # detail that the generic RuntimeError below would hide.
+            import_err = exc
+
+    if not _ops_registered():
+        if import_err is not None:
+            # Surface the raw import error (typically ``ImportError`` with
+            # ``undefined symbol`` ABI detail) instead of burying it in a
+            # generic message — that detail is what tells the user the
+            # mismatch is between libdeepmd_op_pt.so and the runtime torch,
+            # not a missing build.
+            raise import_err
         raise RuntimeError(
             "torch.ops.deepmd_export.{border_op,border_op_backward} "
             "are not registered. Build libdeepmd_op_pt.so and ensure "