test(cuda.core): refactor managed-ops fixtures and consolidate tests

Net -140 LOC across three coupled changes:

* Fixtures 8 -> 5. Drop `memory_pool_*` (consumers re-target to
  `location_ops_*`; every batch test exercises prefetch which needs
  `concurrent_managed_access` anyway), `discard_prefetch_mr`, and the
  two `*_buffer` fixtures.

* Add parametrized `managed_buffer` (pool | external) for prefetch /
  discard / discard_prefetch coverage. Add `external_managed_buffer`
  (non-parametrized) for cuMemAdvise tests — pool-allocated managed
  memory returns NOT_SUPPORTED for some advise values on certain
  driver/device combos, so advise tests stay external-only to match
  the original design's documented carve-out.

* Delete 5 redundant standalone tests; fold 2 (`test_operation_validation`,
  `test_advise_location_validation`) into `TestManagedBuffer`. Keep
  `test_managed_memory_operations_reject_non_managed_allocations`
  standalone — it wraps a device-only pointer, orthogonal to the
  pool/external axis.

* Pull `test_rejects_single_buffer` and `test_length_mismatch` out of
  all three batch classes; replace with two module-level parametrized
  tests covering all three batch fns.

Addresses PR #1775 review comments from leofang (U4, U5, U7).

Author: rparolin