fix(test): reset _V2_BINDINGS cache so legacy-signature tests take the legacy path

The _V2_BINDINGS cache in _buffer.pyx persists across tests, so
monkeypatching get_binding_version alone is insufficient when earlier
tests have already populated the cache with the v2 value. Promote
_V2_BINDINGS from cdef int to a Python-level variable so tests can
monkeypatch it directly via monkeypatch.setattr, and reset it to -1
in both legacy-signature tests.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: rparolin

cuda_core/cuda/core/_memory/_buffer.pyx

@@ -124,7 +124,7 @@ cdef unsigned long long _MANAGED_OPERATION_FLAGS = 0
 # Lazily cached values for immutable runtime properties.
 cdef object _CU_DEVICE_CPU = None
 cdef dict _ADVICE_ENUM_TO_ALIAS = None
-cdef int _V2_BINDINGS = -1
+_V2_BINDINGS = -1
 cdef int _DISCARD_PREFETCH_SUPPORTED = -1
 
 

cuda_core/tests/test_memory.py

@@ -1314,6 +1314,7 @@ def fake_cuMemAdvise(ptr, size, advice, location):
         return (driver.CUresult.CUDA_SUCCESS,)
 
     monkeypatch.setattr(_buffer, "get_binding_version", lambda: _LEGACY_BINDINGS_VERSION)
+    monkeypatch.setattr(_buffer, "_V2_BINDINGS", -1)
     monkeypatch.setattr(_buffer.driver, "cuMemAdvise", fake_cuMemAdvise)
 
     managed_memory.advise(buffer, "set_read_mostly")
@@ -1338,6 +1339,7 @@ def fake_cuMemPrefetchAsync(ptr, size, location, hstream):
         return (driver.CUresult.CUDA_SUCCESS,)
 
     monkeypatch.setattr(_buffer, "get_binding_version", lambda: _LEGACY_BINDINGS_VERSION)
+    monkeypatch.setattr(_buffer, "_V2_BINDINGS", -1)
     monkeypatch.setattr(_buffer.driver, "cuMemPrefetchAsync", fake_cuMemPrefetchAsync)
 
     managed_memory.prefetch(buffer, device, stream=stream)