dials for threshold structure and parallel regression assumption arguments

[corybrunson]

This PR introduces dials for tuning two proposed new ordinal_reg model parameters; see this parsnip PR. Both are specific to ordinal regression.

• threshold_structure controls the constraints on the cut-points between the response value ranges for ordinal regression. It is categorical and takes 4 values, corresponding to specifications shared (under different conventions) by the clm and vglm engines in ordered. (ordered provides an additional option for vglm.) Other than "flexible" (the usual default), it removes one degree of freedom from the model.
• parallel_reg controls whether the parallel regression assumption is imposed on the predictors. While the proposed model argument is quite flexible, the dial is logical; it only supports imposing the assumption on (a) all or (b) no predictors. These are the only two options supported by all three engines that accept the model argument: clm, vglm, and ordinalNet.

As mentioned in the parsnip PR, i'm very open to critical feedback, including on whether these dials are warranted in the first place. I designed them with a brms engine also in mind—the package supports user control over both behaviors—though i since learned that there is some uncertainty over whether it would be appropriate (see the parsnip PR).