fix(cuda.core): allow NUMA pools on older drivers - #2858

Thanks for the fix. The diagnosis in #2795 is right: a driver older than 12.9 cannot answer CU_DEVICE_ATTRIBUTE_HOST_NUMA_MEMORY_POOLS_SUPPORTED, and cuda.core reads the error as False.

Removing the check from allocate() goes too far, though. A 12.9+ driver can answer the query, and a False answer there is real. #2487 added the check because a pool on an unsupported device can return buffers that fail at first use (#2486).

A simpler fix is to repair the property itself. allocate() already reads it, so allocate() then needs no change, and host_numa_memory_pools_supported also stops reporting False on 12.6–12.8 drivers. For example, in _device.pyx (with cy_driver_version cimported from cuda.core._utils.version):

@property
def host_numa_memory_pools_supported(self) -> bool:
    if cy_driver_version() >= (12, 9, 0):
        return bool(
            self._get_cached_attribute(driver.CUdevice_attribute.CU_DEVICE_ATTRIBUTE_HOST_NUMA_MEMORY_POOLS_SUPPORTED)
        )
    # The attribute arrived in CUDA 12.9, HOST_NUMA pools in 12.6 (#2795).
    return cy_driver_version() >= (12, 6, 0) and self.memory_pools_supported and not self.integrated

The integrated clause keeps today's answer on Jetson. The evidence in #2795 comes from a discrete GPU, and #2486 showed host pools on Jetson returning unusable buffers.

Please also:

• Revert the change to allocate().
• Add a test that simulates a pre-12.9 driver.
• Say in the property's docstring that the value is inferred on drivers before 12.9.
• Rebase on main.

When the update is in, I will start the GPU CI.

Actionable comments posted: 1