test(core.utils): drop pre-cache get_kernel in cache roundtrip test

Loading the kernel on the freshly compiled ObjectCode before caching
pre-warms driver-side module state, so a post-reload get_kernel on the
deserialised copy would no longer uniquely prove that the cache flow
works. Keep only the post-reload get_kernel call as the roundtrip
guarantee.

Author: cpcloud

cuda_core/tests/test_program_cache.py

@@ -623,7 +623,9 @@ def test_cache_roundtrip_with_real_compilation(backend, tmp_path, init_cuda):
         compiled = program.compile(target_type)
     finally:
         program.close()
-    assert isinstance(compiled.get_kernel("my_kernel"), Kernel)
+    # Do NOT call get_kernel() on ``compiled`` here: loading the module on the
+    # driver would mutate driver-side state on the original ObjectCode and
+    # weaken the roundtrip guarantee below.
 
     key = make_program_cache_key(
         code=code,