fix(cuda.core): define _read_preferred_location_v2 stub on cu12 builds

Commit 7344d89 lifted the `_read_preferred_location_v2` import in
`_managed_buffer.py` from deferred (inside the property getter) to
module top, per leofang's "no deferred imports" guidance. But that
function was only defined inside the `IF CUDA_CORE_BUILD_MAJOR >= 13`
block in `_managed_memory_ops.pyx`, so on a cu12 build the symbol
doesn't exist and `import cuda.core` fails with

  ImportError: cannot import name '_read_preferred_location_v2' from
  'cuda.core._memory._managed_memory_ops'

Add an ELSE branch defining a stub that raises `NotImplementedError`.
`ManagedBuffer.preferred_location` already gates on both
binding_version() and driver_version() >= (13, 0, 0) before calling,
so the stub is unreachable at runtime — it exists only to satisfy
the unconditional module-level import on cu12.

Fixes CI failure on all 12.9.1 test jobs.

Author: rparolin

cuda_core/cuda/core/_memory/_managed_memory_ops.pyx

@@ -367,6 +367,16 @@ IF CUDA_CORE_BUILD_MAJOR >= 13:
                 loc_arr.data(), loc_indices.data(), <size_t>n,
                 0, hstream,
             ))
+ELSE:
+    def _read_preferred_location_v2(Buffer buf):
+        # Symbol exists so _managed_buffer.py can `from ... import
+        # _read_preferred_location_v2` unconditionally at module top.
+        # `ManagedBuffer.preferred_location` gates on both
+        # binding_version() and driver_version() >= (13, 0, 0) before
+        # calling, so this path is unreachable on a cu12 build.
+        raise NotImplementedError(
+            "_read_preferred_location_v2 requires a CUDA 13 build of cuda.core"
+        )
 
 
 cdef void _do_batch_prefetch(tuple bufs, tuple locs, Stream s):