Improve error handling in SharedTensor.close() and drop()

Changed exception handling in SharedTensor methods:
- close(): Changed from silently logging a warning to properly logging
  an error with the exception message. This provides better visibility
  when shared memory cleanup fails, while still not raising exceptions
  that could cause issues in cleanup paths (like __del__).
- drop(): Fixed exception chaining to use 'raise ... from e' for better
  error traceability and to comply with flake8 B904.

Test plan:
- Ran pytest tests/unit_tests/util/test_shared_tensor.py
- 64 tests passed (1 pre-existing segfault in test_multiprocess_bidirectional)

Author: casteryh

src/forge/util/_shared_tensor.py

@@ -6,18 +6,20 @@
 
 from __future__ import annotations
 
+import logging
+
 import uuid
 from dataclasses import dataclass
 from multiprocessing import shared_memory
 from typing import Optional, Tuple, Union
-import logging
 
 import numpy as np
 import torch
 
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.INFO)
 
+
 @dataclass
 class SharedTensorHandle:
     shm_name: str
@@ -373,8 +375,8 @@ def close(self):
 
         try:
             self._shm.close()
-        except Exception:
-            logger.warn("Error closing shared memory: ", self._shm_name)
+        except Exception as e:
+            logger.error(f"Error closing shared memory {self._shm_name}: {e}")
 
     def drop(self):
         """
@@ -396,8 +398,10 @@ def drop(self):
         # Then unlink
         try:
             self._shm.unlink()
-        except Exception:
-            raise RuntimeError("Error unlinking shared memory: ", self._shm_name)
+        except Exception as e:
+            raise RuntimeError(
+                f"Error unlinking shared memory {self._shm_name}: {e}"
+            ) from e
 
     @property
     def is_closed(self) -> bool: