feat: Complete autodiff for 3 high-priority production layers

Added 4 new TensorOperations and updated 3 production layers:

Operations added:
- ReduceLogVariance: Computes log(variance) along axis with proper gradients
- RBFKernel: Gaussian RBF kernel with gradients for input, centers, and widths
- AffineGrid: Generates sampling grid from affine transformation matrices
- GridSample: Bilinear interpolation sampling with full gradient support

Layers updated:
- LogVarianceLayer: Uses ReduceLogVariance operation
- RBFLayer: Uses RBFKernel operation for Gaussian RBF computations
- SpatialTransformerLayer: Uses AffineGrid + GridSample for spatial transformations

All implementations include production-ready forward and backward passes
with proper gradient computation for training.

Author: claude