Fix tensor_utils.pack() to correctly pack uint64 scalars.

When packing a given NumPy array into a Tensor message, we use the array's dtype to determine which `Packer` to use. For some types (e.g. NumPy strings) we need to determine the base type rather than the concrete type (i.e. `np.str_`) using the `dtype.type` attribute. This can cause some oddities with integer types, e.g:

```
np.array(2**64-1).dtype == np.uint64
np.array(2**64-1).dtype.type == np.ulonglong
np.array(2**64-1 dtype=np.uint64).dtype.type == np.uint64
```

Therefore we re-cast as a dtype (i.e. `dtype(array.dtype.type)`) before looking up the packer.

PiperOrigin-RevId: 410004520
Change-Id: I92cb480bfd0949d021c01d8b41345132cda730dd

Author: tomwardio

dm_env_rpc/v1/tensor_utils.py

@@ -143,7 +143,7 @@ def unpack(self, proto: TensorOrTensorSpecValue) -> np.ndarray:
 }
 
 _TYPE_TO_PACKER = {
-    packer.np_type: packer for packer in _PACKERS
+    np.dtype(packer.np_type): packer for packer in _PACKERS
 }
 
 _DM_ENV_RPC_DTYPE_TO_NUMPY_DTYPE = {
@@ -216,7 +216,7 @@ def get_packer(np_type: np.dtype) -> Packer:
   Raises:
     TypeError: If the provided NumPy type has no known packer.
   """
-  packer = _TYPE_TO_PACKER.get(np_type)
+  packer = _TYPE_TO_PACKER.get(np.dtype(np_type))
   if not packer:
     raise TypeError(f'Unknown NumPy type "{np_type}" has no known packer.')
   return packer

dm_env_rpc/v1/tensor_utils_test.py

@@ -83,6 +83,7 @@ def test_pack_scalar_any_proto(self):
       (25, np.int64, 'int64s'),
       (25, np.uint32, 'uint32s'),
       (25, np.uint64, 'uint64s'),
+      (2**64-1, np.uint64, 'uint64s'),
       (True, np.bool, 'bools'),
       (False, np.bool, 'bools'),
       ('foo', np.str, 'strings'),