Change pack_tensor to do unsafe casting when an explicit dtype is p…

…rovided.

This fixes issues when running with NumPy 2.0.

Previously, we would cast using `same_kind`, which mean't only "safe" casts or casts within a kind, like float64 to float32, were allowed. In NumPy 2.0, the meaning of `same_kind` changed such that unsigned and signed integers are no longer deemed the same.

Considering it was still possible for data loss even when using same_kind casting, we propose changing to use `unsafe` when an explicit dtype is provided, requiring the end user to ensure the value being packed in compatible with the dtype.

PiperOrigin-RevId: 678408939
Change-Id: I1770714fb23335f5ee1a9ae5b7f6263394bb8297

dm_env_rpc/v1/spec_manager_test.py

@@ -102,13 +102,9 @@ def test_pack_wrong_shape_raises_error(self):
       self._spec_manager.pack({'foo': [1, 2]})
 
   def test_pack_wrong_dtype_raises_error(self):
-    with self.assertRaisesRegex(TypeError, 'int32'):
+    with self.assertRaises(ValueError):
       self._spec_manager.pack({'foo': 'hello'})
 
-  def test_pack_cast_float_to_int_raises_error(self):
-    with self.assertRaisesRegex(TypeError, 'int32'):
-      self._spec_manager.pack({'foo': [0.5, 1.0, 1]})
-
   def test_pack_cast_int_to_float_is_ok(self):
     packed = self._spec_manager.pack({'fuzz': [1, 2]})
     self.assertEqual([1.0, 2.0], packed[54].floats.array)

dm_env_rpc/v1/tensor_utils.py

@@ -322,7 +322,7 @@ def pack_tensor(
     value = value.astype(
         dtype=_DM_ENV_RPC_DTYPE_TO_NUMPY_DTYPE.get(dtype, dtype),
         copy=False,
-        casting='same_kind' if value.size else 'unsafe')
+        casting='unsafe')
 
   packed.shape[:] = value.shape
   packer = get_packer(value.dtype.type)

dm_env_rpc/v1/tensor_utils_test.py

@@ -168,7 +168,7 @@ def test_packed_rowmajor(self):
     np.testing.assert_array_equal([1, 2, 3, 4, 5, 6], tensor.int32s.array)
 
   def test_mixed_scalar_types_raises_exception(self):
-    with self.assertRaises(TypeError):
+    with self.assertRaises(ValueError):
       tensor_utils.pack_tensor(['hello!', 75], dtype=np.float32)
 
   def test_jagged_arrays_throw_exceptions(self):