Use run-time array sizing for "work" arrays

[ketch]

In, for instance, the 3D Euler solvers, there are some work arrays whose size is hard-coded. It would be cleaner to size them at run-time, but currently this incurs a performance penalty; see discussion in #75. It seems that in some tests the dynamic allocation leads to a 6-7% slowdown. Once we switch to modules for Riemann solvers, we can size these at run time but still only allocate them once, as module data.