This PR adds optional GPU acceleration to the covariance matrix decomposition
and rotation-to-quaternion conversion, providing significant performance
improvements for CUDA users while maintaining backward compatibility.

## Changes

### `linalg.py`
- Add `use_gpu` parameter to `quaternions_from_rotation_matrices()` (default: True)
- Add pure PyTorch GPU implementation using Shepperd's method
- Original scipy CPU implementation preserved as fallback
- ~300x faster for large batches (2M+ gaussians)

### `gaussians.py`
- Add `use_gpu` parameter to `decompose_covariance_matrices()` (default: True)
- GPU path: SVD on GPU + vectorized reflection correction
- CPU path: original float64 behavior preserved for maximum precision
- Automatic fallback to CPU if GPU SVD fails

## Performance

Tested on RTX 4090 with ~700k gaussians per frame:
- Before: ~4.0s per frame (3s quaternion conversion on CPU)
- After: ~1.0s per frame
- **4x overall speedup**

The bottleneck was `scipy.spatial.transform.Rotation.from_matrix()` which
requires CPU transfer and numpy conversion. The new GPU implementation
stays entirely on device.

## Backward Compatibility

- Default behavior unchanged for CPU tensors
- Set `use_gpu=False` to force original CPU behavior
- API is fully backward compatible (new parameter has default value)