Add OrthogonalRandomFeaturesKernel

[bulochka68]

Summary

Adds OrthogonalRandomFeaturesKernel - a drop-in replacement for RFFKernel
that uses block-orthogonal random projections instead of i.i.d. Gaussian samples.

ORF strictly dominates standard RFF in mean squared error of kernel approximation
(Choromanski et al., 2017) while maintaining the same computational complexity O(nD)
and the same interface.

Usage

# Before
kernel = gpytorch.kernels.RFFKernel(num_samples=512)

# After — identical interface, better approximation
kernel = gpytorch.kernels.OrthogonalRandomFeaturesKernel(num_samples=512)

Implementation details

Each block of d frequency vectors shares a Haar-distributed orthogonal basis (via QR
decomposition) but has independent chi(d)-distributed norms, so the marginal distribution
of each frequency matches N(0, I_d) — preserving the unbiasedness of the RBF approximation.

Changes

• gpytorch/kernels/orf_kernel.py — new kernel
• gpytorch/kernels/__init__.py — export
• test/kernels/test_orf_kernel.py — 18 tests (BaseKernelTestCase + ORF-specific)
• docs/source/kernels.rst — documentation entry

References

• Yu et al., Orthogonal Random Features, NeurIPS 2016
• Choromanski et al., The Geometry of Random Features, AISTATS 2018