Spectral time kernel

[roussel-ryan]

• implements a spectral mixture kernel for the time axis of time dependent model constructor, inspired by the work of @nikitakuklev
• spectral kernel is now the default for TimeDependentModelConstructor
• behavior can be disabled by setting use_spectral_mixture_kernel=False

[nikitakuklev]

GPyTorch recommends to initialize SMK from data. For example, from code in the paper above:

    if enable_time:
        smk = SpectralMixtureKernel(num_mixtures=num_smk_mixtures, ard_num_dims=1, active_dims=(dx.shape[-1] - 1,))
        smk.initialize_from_data(dx, dy)

        covar_module = ProductKernel(smk, regular_kernel)
        subset_batch_dict = {
            "likelihood.noise_covar.raw_noise": -2,
            "covar_module.base_kernel.kernels.1.raw_outputscale": -1,
            "covar_module.base_kernel.kernels.1.base_kernel.raw_lengthscale": -3,
        }

The subset dict part is not used by us, but initialization should be helpful. Looking at source of that heuristic, it has some statistical tests and shouldn't be a performance problem. So, maybe just do it here as well?

[nikitakuklev]

Also, I found 3 components a more stable choice generally.

[roussel-ryan]

@nikitakuklev thanks for your comments, lmk if these changes address them properly

[nikitakuklev]

Good idea to leave init as choice. LGTM.