Conditional solver selection

[AdrianSosic]

Transfer learning using the scipy solver invoked via fit_gpytorch_mll is currently unstable because it seems we run into overfitting problems due to missing priors on the involved IndexKernel. Using the old fit_gpytorch_mll_torch routine mitigates this problem, probably because the MLL fitting does not fully converge, which effectively acts as regularization via early stopping. In a non-transfer-learning context, however, the scipy solver produces generally better results.

This PR thus implements a temporary workaround where the solver is chosen adaptively based on the presence of a TaskParameter.