Fix tests and code before CI implementation

[emapuljak]

Today tests/run_tests.sh is the only place that knows which test module runs under which
backend, so a plain pytest collects torch-only modules under the keras backend and fails at import
(cannot import name 'PQBatchNorm2d' from 'pquant.core.keras.layers'). CI's matrix runs plain
pytest per backend, so this must be fixed in the test files themselves.

Implementation note: gating must use pytest.skip(..., allow_module_level=True) placed
above the backend-specific imports (or pytest.importorskip). A pytest.mark.skipif or a
custom marker will not work — markers are evaluated after the module is imported, and the
failure here happens at import time. (This is also why no markers = [...] entry is needed in
[tool.pytest.ini_options].)

• Backend-gate each backend-specific test module so it skips at collection instead of
  erroring. Guard before the backend-specific imports, e.g.:
  python import keras, pytest if keras.backend.backend() != "torch": pytest.skip("torch backend only", allow_module_level=True)
  Apply to the torch-only modules (test_torch_compression_layers.py,
  test_torch_onnx_converter.py, test_torch_alkaid_conversion.py, test_hgq_torch.py) and
  symmetrically to the keras-only ones (test_keras_*, test_hgq_keras.py). Leave
  test_ap.py / test_pdp.py / test_wanda.py ungated — they run under both.
• Remove the ineffective os.environ["KERAS_BACKEND"] = "torch" at the top of
  test_torch_compression_layers.py — it's a no-op once keras is already imported by
  conftest.py, and it masks the real gating problem.
• Skip optional-dependency tests when the dep is absent: add pytest.importorskip("onnx")
  to the onnx converter tests and pytest.importorskip("alkaid") to the alkaid conversion
  tests. (alkaid is the plugin host — it must not become a pquant dependency; onnx is
  opt-in via the [onnx] extra.)
• Preserve channels_last coverage: run_tests.sh runs test_keras_compression_layers.py
  a second time with DATA_FORMAT=channels_last. Keep this axis alive by parametrizing the
  set_image_data_format fixture in conftest.py over channels_first/channels_last (or by
  adding a dedicated CI job) — otherwise it silently stops being tested.
• Delete tests/run_tests.sh once the above lands — KERAS_BACKEND=<backend> pytest then
  selects the right tests automatically, and the CI matrix supersedes the script.
• Fix CUDA-forcing in the torch tests so they run on CPU-only machines. ~40 torch tests
  currently fail with AssertionError: Torch not compiled with CUDA enabled
  (torch/cuda/__init__.py:522) on any machine without a GPU. conftest.py already tries to
  fall back correctly (device = "cuda" if torch.cuda.is_available() else "cpu"), so something
  in the tests or library code is initialising CUDA regardless — find and fix it. This is a CI
  blocker: GitHub's ubuntu-latest runners have no GPU, so these tests will fail in the
  KERAS_BACKEND=torch matrix jobs exactly as they do locally on macOS. Verify with
  KERAS_BACKEND=torch pytest tests on a CPU-only machine — expect zero CUDA-related failures.
• Fix hardcoded CUDA in add_compression_layers. src/pquant/core/torch/layers.py:641-644
  calls model.to("cuda") (and .to("cuda") on the warm-up tensor) unconditionally, with no
  torch.cuda.is_available() guard — so the function crashes with
  AssertionError: Torch not compiled with CUDA enabled on any CPU-only machine. This is the
  root cause of the ~40 failing torch tests, it makes the library unusable for CPU users, and it
  blocks the CI torch matrix (GitHub runners have no GPU). Fix by resolving the device once
  ("cuda" if torch.cuda.is_available() else "cpu") and using it for all three calls. Verify
  with KERAS_BACKEND=torch pytest tests on a CPU-only machine.

[emapuljak]

Fix hardcoded CUDA is fixed with #59