PointerType should be rendered differently in OpenCL backend

This fixes test for LLVM GPU using pointer_arith_func

Using the SymPy implementation for printing a `sympy.Function` has some issues:

 - you cannot use `cast_func` as an argument -> SymPy prints
 `cast_func(2,float)
 - typed numbers are not typed any more -> e.g. pow(x, 0.25) instead of
 pow(x, 0.25f)

- bugfix: loop counter of vectorized loop now correctly stored as SIMD vector with entries i, i+1, i+2, ...
- basis for in-kernel boundary handling

For sympy, oo s a number. So pystencils prints a type infinity a double
INFINITY as INFINITY.0

Pystencils will print two const when printing an assignment to a const numpy.dtype.
With the new type interference, this case can now occur more often

There is now a superior implementation in pystencils_autodiff.

This is actually a AST Node for wrapper code. pystencils_autodiff has
now a AST printer for framework integration code. So framework AST nodes
are now there and kernel AST nodes remain here.

headers and global_declarations are generated by methods that return
sets. So even with the same inputs it is not guaranteed that the same
source code is generated since sets do not guarantee a specific order
when iterating over them.

Sum and Product have a indexing variable which is a Atom but not a free
symbol

- for example: sqrt(cast_func(3, double)) previously printed "sqrt(3)"
  instead of "sqrt(3.0)" leading to errors in CUDA code

`instruction_set`

This commit avoid dependency of core pystencils on jinja2.
However this could make the printing of some AST-nodes less elegant.