coverage: add cuda.core tests for system device, checkpoint, memory, stream, green context, and tensor map

[rluo8]

Description

Adds focused unit tests across cuda.core, targeting input-validation branches, error paths, and object-lifecycle behavior that were previously uncovered. All tests are real-path: host-only tests exercise pure helpers / argument validation directly, and GPU-required tests use the existing fixtures and skips. No production changes.

Measured effect on cuda/core/ line coverage : 84.34% → ~85.6%, with system/_device.pyx crossing 85%. Largest gains in checkpoint.py, _stream.pyx, _memory/_memory_pool.pyx, _tensor_map.pyx, andsystem/_device.pyx.
Note that the coverage numbers are for cuda/core/ only, combined across the local systems Linux (A10), Windows (A10), and Linux (H200).

File	Tests
graph/test_graph_definition.py	host-callback node reconstruction via _create_from_driver through an embedded child graph
system/test_system_device.py	temperature / constructor-selector / affinity·topology·P2P argument validation (rejected before any NVML call)
test_checkpoint.py	TestCheckpointHelpers (host-only _check_timeout_ms / _make_restore_args / _as_cuuuid marshalling + validation) and TestCheckpointDriverDispatch (_call_driver result-code / missing-symbol translation via a dependency-injected boundary fake)
test_green_context.py	closed-context handle/is_green/resources degradation, empty/wrong-type count split rejection, primary-context stream SM-resource path
test_launcher.py	test_launch_scalar_argument_subclass_fallback expanded to all ctypes scalar widths/signs + numpy, with kernel read-back asserting correct C-type marshalling (guards e.g. uint64 → uint32_t truncation)
test_memory.py	mempool DeviceMemoryResource handle ownership, idempotent close, deallocate-frees-pointer (verified via used_mem_current drop, not just handle == 0), IPC-unsupported raise, _MemPoolAttributes non-instantiable
test_stream.py	__cuda_stream__ protocol validation (malformed value / wrong length / bad version / deprecated attribute form), borrow-from-existing-Stream, from_handle flag/priority query, per-thread vs legacy default-stream env selection
test_tensor_map.py	TMA descriptor arg rejection branches — box_dim/enum via the public TensorMapDescriptorOptions(...) constructor, dtype and box_dim-required via the host-only marshalling helpers

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2404/
https://nvidia.github.io/cuda-python/pr-preview/pr-2404/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2404/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2404/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

The system tests look good to me. This should probably get reviewed by specialists in other areas of the code as well.