[CI/Build] Fix AMD CI: test_cpu_gpu.py

[zhewenl]

Purpose

test_cpu_gpu.py is added in #21448, while it add different attention backends to test, some backends might not be compatible with platforms - like FlashInferBackend is not supported in ROCM at the current moment.

This PR refactors the test to conditional import backends, like what we did in test_attention_backends.py.

Test Plan

pytest -v -s v1/kv_offload/test_cpu_gpu.py
INFO 10-22 21:15:25 [__init__.py:225] Automatically detected platform rocm.
================================================================= test session starts =================================================================
platform linux -- Python 3.12.11, pytest-8.4.2, pluggy-1.6.0 -- /home/zhewenli/uv_env/vllm-fork/bin/python3
cachedir: .pytest_cache
rootdir: /data/users/zhewenli/gitrepos/vllm-fork
configfile: pyproject.toml
plugins: anyio-4.11.0, asyncio-1.2.0
asyncio: mode=Mode.STRICT, debug=False, asyncio_default_fixture_loop_scope=None, asyncio_default_test_loop_scope=function
collecting ... WARNING 10-22 21:15:43 [interface.py:512] Current platform cuda does not have '__test__' attribute.
WARNING 10-22 21:15:43 [interface.py:512] Current platform cuda does not have '__bases__' attribute.
WARNING 10-22 21:15:43 [interface.py:512] Current platform cuda does not have '__test__' attribute.
collected 4 items                                                                                                                                     

v1/kv_offload/test_cpu_gpu.py::test_transfer[cuda:0-0-dtype0-4-256-64-1-16-8-64-3-True] INFO 10-22 21:15:43 [cpu_gpu.py:78] Allocating 4 CPU tensors...
PASSED
v1/kv_offload/test_cpu_gpu.py::test_transfer[cuda:0-0-dtype0-4-256-64-1-16-8-64-3-False] INFO 10-22 21:15:44 [cpu_gpu.py:78] Allocating 4 CPU tensors...
PASSED
v1/kv_offload/test_cpu_gpu.py::test_transfer[cuda:0-0-dtype0-4-256-64-3-16-8-64-3-True] INFO 10-22 21:15:44 [cpu_gpu.py:78] Allocating 4 CPU tensors...
PASSED
v1/kv_offload/test_cpu_gpu.py::test_transfer[cuda:0-0-dtype0-4-256-64-3-16-8-64-3-False] INFO 10-22 21:15:45 [cpu_gpu.py:78] Allocating 4 CPU tensors...
PASSED

================================================================== warnings summary ===================================================================
<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
============================================================ 4 passed, 2 warnings in 2.86s ============================================================