speed up a brutal macro-expansion torture test by about 30% (1.5 -> 1.0s)

by turning vectors of vectors into a single vector, reducing pressure on
malloc.  This can still be improved.

llvm-svn: 38753