switch from model.matrix to sparseMMFromMatrix

[josherrickson]

See comments on 283e156.

To replace mlm.R:101-102's use of model.matrix with SparseMMFromMatrix, which currently lives in the clusters branch of RItools.

The following code almost works, though nt or nc will have the wrong length if the optmatch object is manually manipulated (e.g. this test). I believe there's a bug in either SparseMMFromMatrix or something with matrix.csr ops, which I will crosspost to RItools about.

sparseMM <- SparseMMFromFactor(theMatch)
nt.csr <- t(as.matrix.csr(z))%*%sparseMM
nt <- attr(nt.csr, "ra")[order(attr(nt.csr, "ja"))]
nc.csr <- t(as.matrix.csr(!z))%*%sparseMM
nc <- attr(nc.csr, "ra")[order(attr(nc.csr, "ja"))]

[josherrickson]

I ended up copying SparseMMFromFactor from RItools (after all, this is an experimental repo) so I could implement the changes in d0e7fe4. I see another big performance gain; the example in tests/mlm_time_test.R now takes ~.03sec to run, down from the previous ~.2 and the original ~4.

When we move mlm to optmatch, we may need to change the imports a bit to point to the proper RItools location, so leaving the issue open.