Multi-channel audio is automatically cast to mono, num_channels is ignored

### Describe the bug

The `num_channels` parameter in `datasets.Audio()` is documented to preserve stereo channels when set to `None` (preserve original) or `2` (explicit stereo), but it currently downmixes all audio to mono regardless of this setting.

### Steps to reproduce the bug

```python
import numpy as np
import soundfile as sf
import tempfile
from datasets import Dataset, Audio

# Create a stereo audio file
sample_rate = 16000
duration = 1.0
num_samples = int(sample_rate * duration)

left_channel = np.sin(2 * np.pi * 440 * np.linspace(0, duration, num_samples))
right_channel = np.sin(2 * np.pi * 880 * np.linspace(0, duration, num_samples))
stereo_audio = np.stack([left_channel, right_channel], axis=1).astype(np.float32)

temp_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
sf.write(temp_file.name, stereo_audio, sample_rate)

# Create HuggingFace dataset
dataset_dict = {"audio": [temp_file.name]}
ds = Dataset.from_dict(dataset_dict)

# Test with num_channels=2
ds_stereo = ds.cast_column("audio", Audio(num_channels=2))
audio_data = ds_stereo[0]["audio"]

print(f"Original file shape (via soundfile): {sf.read(temp_file.name)[0].shape}")
# Output: (16000, 2) ✓ Stereo

print(f"HF datasets shape with num_channels=2: {audio_data['array'].shape}")
# Output: (16000,) ✗ Mono (should be (2, 16000))
```

**Result:**
- Original file: `(16000, 2)` - stereo ✓
- `Audio(num_channels=None)`: `(16000,)` - mono ✗
- `Audio(num_channels=2)`: `(16000,)` - mono ✗
- `Audio(num_channels=1)`: `(16000,)` - mono ✓

### Expected behavior

According to the documentation, `Audio` decoding should return samples with shape `(num_channels, num_samples)`:
- `num_channels=None` should preserve the original number of channels from the source file
- `num_channels=2` should preserve/convert to stereo output with shape `(2, num_samples)`
- `num_channels=1` should downmix to mono with shape `(num_samples,)`

**Actual Behavior**
All `num_channels` settings produce mono output with shape `(num_samples,)`, even when the source audio file is stereo.

### Environment info

OS: macOS / Linux
Python 3.10.19
```
datasets 4.4.2
torchcodec 0.10.0
```