Apparently string slices are not as smart as I was hoping for. Slicing a looong (say, 1MB) string and holding onto it is not a good idea if done thousands of times. So, do fewer slicings and decompile subtables immediately instead of holding onto data. This makes me want to rethink the kind of data structures we use for lazy processing. Fixes https://github.com/behdad/fonttools/issues/317