Only refactor matrices when necessary

[JHopeCollins]

Context
The default behaviour of LinearVariationalSolver is to reassemble the matrix and preconditioner at every solve call. If the matrix type is not mat_free this means recalculating all the entries of the matrix, and if a factorisation method is used for the preconditioner (e.g. lu, ilu) then this will be refactorised.

Problem
If the matrix hasn't actually changed then reassembly is unnecessary. Mass matrix solvers for explicit schemes are the most obvious example of this, but anywhere where a constant reference profile is used will also be doing this.
These operations can be quite expensive. In particular, factorisations do not scale well for large problems or core counts.

Solution
LinearVariationalProblem has a constant_jacobian argument to the constructor. If True then the matrix will not be factorised at every step.

For #542, reassembly will only be required when the reference profiles are updated. The constant_jacobian argument can still be set, and the LinearVariationalSolver.invalidate_jacobian() can be called whenever the profiles are updated to tell the solver to reassemble at the next solve call.

Related, sometimes there might be times when a NonlinearVariationalSolver is actually solving a linear system that doesn't need reassembling. If that happens, then these PETSc options will prevent reassembly after the first solve:

'snes_lag_jacobian': -2,
'snes_lag_jacobian_persists': None,
'snes_lag_preconditioner': -2,
'snes_lag_preconditioner_persists': None