In the end I had to combine the last two versions. For jacobians:
- We want to split according to the FunctionView (e.g. Stokes).
- We don't want to split according to gradients (e.g. sumfactorization).

Example: Before this commit the expression

\sum_i a_i (\nabla v)_i + \sum_j b_j (\nabla v)_j

was split into two accumulation terms:

1) a_k with corresponding test function (\nabla v)_k
2) b_l with corresponding test function (\nabla v)_l

Now we split into:

a_k + b_k with corresponding test function (\nabla v)_k

This is possible since we have linearity in the test function.

TODO: Jacobians are not yet working

The backend selecting mechanism needs all modules to be imported.
This commit restructures the package to be able to do that. Placement
of some functions felt weird, but this can still be changed.
No more placements in __init__.py anymore though, as this will
result in cyclic dependencies.

including CallInstructions for accumulation etc.

laplace is again proceeding to code generation

UFL did not like  my way of replacing some args with 1 as it
destroyed the shape information. It now remains in the expression,
but triggers code generation separately.

and remove old implementation