[FEA] cudf-polars test suite CLI arguments for `NUM_RANKS` and `allow_gpu_sharing`

`python/cudf_polars/tests/conftest.py` currently hardcodes both the rank count and the GPU-sharing flag for multi-rank streaming engines:

```python
# Number of ranks for multi-rank streaming engines that share one GPU
# (currently ``RayEngine``). Single-GPU dev hosts and CI runners require
# ``allow_gpu_sharing=True`` to oversubscribe one device across actors.
NUM_RANKS = 2
```

…and:

```python
engines["ray"] = RayEngine(
num_ranks=NUM_RANKS,
engine_options={"allow_gpu_sharing": True},
ray_init_options={"include_dashboard": False},
)
```

This configuration works for the initial single-GPU CI lane, but it hardcodes assumptions that belong to the test runner environment rather than library-level test setup.

### Proposed

Make both `NUM_RANKS` and `allow_gpu_sharing` configurable from outside `conftest.py`, while preserving the current behavior as the default.

Register pytest CLI options via `pytest_addoption`, expose them as session-scoped fixtures, and consume them from `streaming_engines`:

```bash
pytest --num-ranks=4 --no-gpu-sharing ...
```