Support sparse output for sandwich products

[ElizabethSantorellaQC]

There are situations where we can expect ahead of time that the output of a sandwich product will be sparse:

• If the data is dominated by one very high-dimensional categorical. For example, say there are D dense columns, and M categorical columns. The fraction of nonzeros in the resulting sandwich product will be at most (D^2 + 2 DM + M) / (D + M)^2. If M >> D this matrix will be quite sparse. This could happen in an e-commerce pricing context, if features are a categorical product ID that is very high-dimensional and a small number of scalars.
• Perhaps we have computed x.sandwich(d1) and now we want to know x.sandwich(d2). The latter will have the same sparsity pattern as the former.
• We can analytically upper-bound the number of nonzero elements in a sandwich product of categoricals using the number of rows. If the data is made up mainly of categoricals and is very wide relative to its length, its sandwich product will be fairly sparse. This would be true in, for example, the German employer-employee data set used in the AKM papers. Since a typical worker will usually not have worked for a randomly-chosen firm, the sandwich product would be very sparse.