cuda.core: VirtualMemoryResourceOptions accepts values the driver rejects, and modify_allocation mishandles edge inputs

Validation gaps found while analyzing #2882, each a small change: (1) addr_align larger than the granularity makes the adjacent-reservation probe fail with a misaligned hint, which is treated as fatal instead of falling back to the slow path; (2) new_size < buf.size is silently treated as an access-only update; (3) allocation_type=MANAGED is exposed but current drivers reject it in cuMemCreate; (4) modify_allocation guards a user-reachable condition with assert; (5) allocate(0) reaches the driver and fails while sibling resources return an empty Buffer; (6) the default handle_type=POSIX_FD fails on native Windows WDDM devices and on location_type="host"; (7) for host locations device_id is validated and ignored, HOST_NUMA_CURRENT can never allocate, and a host resource with peers reports is_device_accessible=False. Related: #2694.

Start by tracing the cuda.core VirtualMemoryResourceOptions, modify_allocation, and allocate entry points, then compare their behavior with sibling resources. Check each of the seven listed edge cases, including host locations and Windows WDDM, and add or update coverage so accepted inputs match driver behavior and zero-size allocation matches sibling resources.