Implement an option for factor-based prediction

Right now we can do prediction (i.e., design matrix generation) by specifying a set of data values ({"x": 1, "y": 2}). But for some purposes, like generating all pairwise categorical significance tests ("is the value of my linear predictor significantly different at a == a1 than it is at a == a2?"), it would be nice to be able to write this in terms of factor values ("is the value of my linear predictor significantly different when the C(...) factor is on level a1 compared to when it's on level a2?", without having to grovel through the AST of the factor's python code to figure out which data variables are being accessed. In particular, this would be expressed at the same representational level as factor-level metadata (see #61).

So we should have a version of predict that takes a set of specifications for factor values, and goes from there.

Start by reviewing the existing prediction and design-matrix APIs, then compare their representation with the factor-level metadata discussed in issue #61. Define how factor-value specifications should map to prediction inputs and what pairwise categorical comparisons must produce; the issue names no files or tests to run.