CI: All linux-aarch64 configurations with CTK 13.x fail the TestPrefetchBatch.test_per_buffer_location test

[mdboom]

This breakage happened externally to any change in the codebase. This run worked yesterday, but re-running it today causes it to fail: https://github.com/NVIDIA/cuda-python/actions/runs/28138378349

It is also failing on most open PRs today.

__________________ TestPrefetchBatch.test_per_buffer_location __________________

self = <test_managed_ops.TestPrefetchBatch object at 0xf0e37fc9b880>
location_ops_device = <Device 0 (NVIDIA L4)>
location_ops_mr = <cuda.core._memory._managed_memory_resource.ManagedMemoryResource object at 0xf0e375abf340>

    def test_per_buffer_location(self, location_ops_device, location_ops_mr):
        from cuda.core.utils import prefetch_batch

        device = location_ops_device
        bufs = [location_ops_mr.allocate(_MANAGED_TEST_ALLOCATION_SIZE, stream=device.default_stream) for _ in range(2)]
        stream = device.create_stream()

        prefetch_batch(stream, bufs, [Host(), device])
        stream.sync()

        last0 = _last_prefetch_location(bufs[0])
        last1 = _last_prefetch_location(bufs[1])
>       assert last0 == _HOST_LOCATION_ID
E       assert 0 == -1

bufs       = [<Buffer ptr=0x1910200000 size=4096>, <Buffer ptr=0x1910201000 size=4096>]
device     = <Device 0 (NVIDIA L4)>
last0      = 0
last1      = 0
location_ops_device = <Device 0 (NVIDIA L4)>
location_ops_mr = <cuda.core._memory._managed_memory_resource.ManagedMemoryResource object at 0xf0e375abf340>
prefetch_batch = <cyfunction prefetch_batch at 0xf0e39d903f40>
self       = <test_managed_ops.TestPrefetchBatch object at 0xf0e37fc9b880>
stream     = <Stream handle=0xc60a5a102770 context=0xc60a5789cb50>

tests/memory/test_managed_ops.py:226: AssertionError

[leofang]

This should be fixed by #2268. What happened yesterday was that all linux-aarch64 runners have a 4K page size on host but the size in the VM was bumped to 64K. Though before 2268 landed the size bump was reverted, so the verification will have to wait until it's bumped next time.