Improving PETSc SNES support

At present the current implementation of PETScNonLinearSolver is not fully optimized. These are the main problems.

* [ ] Residuals and jacobians are copied back and forth from PArrays to PETSc objects, and vice-versa. I guess that this can be overcomed by directly assembling into PETSc data structures (as we did in GridapDistributed.jl 0.1.0). I would consider this in a future PR if and only if we determine that it becomes a bottleneck/hot spot.
* [x] The jacobian copy overhead is even more severe because the local storage format of PSparseMatrix by default is not SparseMatrixCSR{0,PetscInt,PetscReal} but SparseMatrixCSC, so that we need a conversion at each Jacobian calculation. I guess this latter issue can be easily overcomed. https://github.com/gridap/GridapPETSc.jl/pull/66/
* [ ] The PETScNonLinearSolverCache does not currently store a reference to the KSP/PC objects used for the previous nonlinear solve. I guess it would be convenient to have this in the future. (We can put it an issue).

Besides,
* Partitioned simulations in sequential mode currently trigger a @notimplemented macro. I did not have enough strength to fully implement this case.