Support precisions other than double

[dham]

Currently we only support double precision real and complex. PETSc supports other precisions.

The general pattern would be to follow what we do for complex.

The pathway to fix would be as follows:

1. Add some sort of precision argument to the installer.
2. Follow through the installer to:
   a. set the scalar type passed to PETSc
   b. ensure that the logic for forcing recompilation when the scalar type changes also applies in this case:
   c. remove any downstream PETSc dependencies that fail at the new precision (this will be experimentally determined).
3. Run our test suite and hunt down the failures. This is likely to be due to hand-coded kernels for things like utility meshes.

[wence-]

Run our test suite and hunt down the failures. This is likely to be due to hand-coded kernels for things like utility meshes.

Note is very likely that many tests will fail due to tolerances that were plucked out of the air assuming double precision as well.